pypocketfft.cc 20 KB
Newer Older
Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
1 2 3 4 5 6
/*
 * This file is part of pocketfft.
 * Licensed under a 3-clause BSD style license - see LICENSE.md
 */

/*
7
 *  Python interface.
Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
8
 *
9
 *  Copyright (C) 2019 Max-Planck-Society
Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
10 11 12
 *  \author Martin Reinecke
 */

Martin Reinecke's avatar
Martin Reinecke committed
13 14
#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>
Martin Reinecke's avatar
more  
Martin Reinecke committed
15
#include <pybind11/stl.h>
Martin Reinecke's avatar
Martin Reinecke committed
16

17
#include "pocketfft_hdronly.h"
Martin Reinecke's avatar
Martin Reinecke committed
18

19
namespace {
20

21
using namespace std;
Martin Reinecke's avatar
Martin Reinecke committed
22 23
using pocketfft::shape_t;
using pocketfft::stride_t;
Martin Reinecke's avatar
Martin Reinecke committed
24

Martin Reinecke's avatar
Martin Reinecke committed
25 26
namespace py = pybind11;

27 28 29 30
// 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;

31 32 33 34 35 36
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;
Martin Reinecke's avatar
Martin Reinecke committed
37
auto None = py::none();
Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
38 39

shape_t copy_shape(const py::array &arr)
Martin Reinecke's avatar
Martin Reinecke committed
40
  {
41
  shape_t res(size_t(arr.ndim()));
Martin Reinecke's avatar
Martin Reinecke committed
42
  for (size_t i=0; i<res.size(); ++i)
43
    res[i] = size_t(arr.shape(int(i)));
Martin Reinecke's avatar
Martin Reinecke committed
44 45
  return res;
  }
Martin Reinecke's avatar
Martin Reinecke committed
46

Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
47
stride_t copy_strides(const py::array &arr)
Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
48
  {
49
  stride_t res(size_t(arr.ndim()));
Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
50
  for (size_t i=0; i<res.size(); ++i)
51
    res[i] = arr.strides(int(i));
Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
52 53
  return res;
  }
Martin Reinecke's avatar
Martin Reinecke committed
54

Martin Reinecke's avatar
Martin Reinecke committed
55
shape_t makeaxes(const py::array &in, const py::object &axes)
Martin Reinecke's avatar
more  
Martin Reinecke committed
56
  {
57
  if (axes.is_none())
Martin Reinecke's avatar
more  
Martin Reinecke committed
58
    {
59
    shape_t res(size_t(in.ndim()));
Martin Reinecke's avatar
Martin Reinecke committed
60 61 62
    for (size_t i=0; i<res.size(); ++i)
      res[i]=i;
    return res;
Martin Reinecke's avatar
more  
Martin Reinecke committed
63
    }
64 65 66
  auto tmp=axes.cast<std::vector<ptrdiff_t>>();
  auto ndim = in.ndim();
  if ((tmp.size()>size_t(ndim)) || (tmp.size()==0))
Martin Reinecke's avatar
more  
Martin Reinecke committed
67
    throw runtime_error("bad axes argument");
68 69 70 71 72 73 74 75
  for (auto& sz: tmp)
    {
    if (sz<0)
      sz += ndim;
    if ((sz>=ndim) || (sz<0))
      throw invalid_argument("axes exceeds dimensionality of output");
    }
  return shape_t(tmp.begin(), tmp.end());
Martin Reinecke's avatar
more  
Martin Reinecke committed
76 77
  }

78
#define DISPATCH(arr, T1, T2, T3, func, args) \
79
  { \
80 81 82
  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; \
83 84
  throw runtime_error("unsupported data type"); \
  }
85

86 87 88 89 90 91 92 93 94
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 invalid_argument("invalid value for inorm (must be 0, 1, or 2)");
  }

template<typename T> T norm_fct(int inorm, const shape_t &shape,
Martin Reinecke's avatar
Martin Reinecke committed
95
  const shape_t &axes, size_t fct=1, int delta=0)
96 97 98 99
  {
  if (inorm==0) return T(1);
  size_t N(1);
  for (auto a: axes)
100
    N *= fct * size_t(int64_t(shape[a])+delta);
101 102 103
  return norm_fct<T>(inorm, N);
  }

Martin Reinecke's avatar
Martin Reinecke committed
104 105 106
template<typename T> py::array_t<T> prepare_output(py::object &out_,
  shape_t &dims)
  {
107
  if (out_.is_none()) return py::array_t<T>(dims);
Martin Reinecke's avatar
Martin Reinecke committed
108 109 110 111 112 113
  auto tmp = out_.cast<py::array_t<T>>();
  if (!tmp.is(out_)) // a new object was created during casting
    throw runtime_error("unexpected data type for output array");
  return tmp;
  }

Martin Reinecke's avatar
Martin Reinecke committed
114
template<typename T> py::array c2c_internal(const py::array &in,
Martin Reinecke's avatar
Martin Reinecke committed
115 116
  const py::object &axes_, bool forward, int inorm, py::object &out_,
  size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
117
  {
Martin Reinecke's avatar
Martin Reinecke committed
118
  auto axes = makeaxes(in, axes_);
Martin Reinecke's avatar
Martin Reinecke committed
119
  auto dims(copy_shape(in));
Martin Reinecke's avatar
Martin Reinecke committed
120
  auto res = prepare_output<complex<T>>(out_, dims);
Martin Reinecke's avatar
Martin Reinecke committed
121 122 123 124 125 126
  auto s_in=copy_strides(in);
  auto s_out=copy_strides(res);
  auto d_in=reinterpret_cast<const complex<T> *>(in.data());
  auto d_out=reinterpret_cast<complex<T> *>(res.mutable_data());
  {
  py::gil_scoped_release release;
127
  T fct = norm_fct<T>(inorm, dims, axes);
Martin Reinecke's avatar
Martin Reinecke committed
128
  pocketfft::c2c(dims, s_in, s_out, axes, forward, d_in, d_out, fct, nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
129
  }
Martin Reinecke's avatar
Martin Reinecke committed
130 131
  return res;
  }
Martin Reinecke's avatar
Martin Reinecke committed
132

Martin Reinecke's avatar
Martin Reinecke committed
133
template<typename T> py::array c2c_sym_internal(const py::array &in,
Martin Reinecke's avatar
Martin Reinecke committed
134 135
  const py::object &axes_, bool forward, int inorm, py::object &out_,
  size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
136
  {
Martin Reinecke's avatar
Martin Reinecke committed
137
  auto axes = makeaxes(in, axes_);
Martin Reinecke's avatar
Martin Reinecke committed
138
  auto dims(copy_shape(in));
Martin Reinecke's avatar
Martin Reinecke committed
139
  auto res = prepare_output<complex<T>>(out_, dims);
Martin Reinecke's avatar
Martin Reinecke committed
140 141 142 143 144 145 146
  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<complex<T> *>(res.mutable_data());
  {
  py::gil_scoped_release release;
  T fct = norm_fct<T>(inorm, dims, axes);
Martin Reinecke's avatar
Martin Reinecke committed
147
  pocketfft::r2c(dims, s_in, s_out, axes, forward, d_in, d_out, fct, nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
148 149 150
  // now fill in second half
  using namespace pocketfft::detail;
  ndarr<complex<T>> ares(res.mutable_data(), dims, s_out);
151
  rev_iter iter(ares, axes);
Martin Reinecke's avatar
Martin Reinecke committed
152 153
  while(iter.remaining()>0)
    {
154 155
    auto v = ares[iter.ofs()];
    ares[iter.rev_ofs()] = conj(v);
Martin Reinecke's avatar
Martin Reinecke committed
156 157 158 159 160 161
    iter.advance();
    }
  }
  return res;
  }

Martin Reinecke's avatar
Martin Reinecke committed
162 163
py::array c2c(const py::array &a, const py::object &axes_, bool forward,
  int inorm, py::object &out_, size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
164
  {
165
  if (a.dtype().kind() == 'c')
Martin Reinecke's avatar
Martin Reinecke committed
166 167
    DISPATCH(a, c128, c64, clong, c2c_internal, (a, axes_, forward,
             inorm, out_, nthreads))
168

Martin Reinecke's avatar
Martin Reinecke committed
169 170
  DISPATCH(a, f64, f32, flong, c2c_sym_internal, (a, axes_, forward,
           inorm, out_, nthreads))
Martin Reinecke's avatar
Martin Reinecke committed
171
  }
172

Martin Reinecke's avatar
Martin Reinecke committed
173
template<typename T> py::array r2c_internal(const py::array &in,
Martin Reinecke's avatar
Martin Reinecke committed
174 175
  const py::object &axes_, bool forward, int inorm, py::object &out_,
  size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
176
  {
Martin Reinecke's avatar
Martin Reinecke committed
177
  auto axes = makeaxes(in, axes_);
Martin Reinecke's avatar
Martin Reinecke committed
178
  auto dims_in(copy_shape(in)), dims_out(dims_in);
Martin Reinecke's avatar
more  
Martin Reinecke committed
179
  dims_out[axes.back()] = (dims_out[axes.back()]>>1)+1;
Martin Reinecke's avatar
Martin Reinecke committed
180
  py::array res = prepare_output<complex<T>>(out_, dims_out);
Martin Reinecke's avatar
Martin Reinecke committed
181 182 183 184 185 186
  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<complex<T> *>(res.mutable_data());
  {
  py::gil_scoped_release release;
187
  T fct = norm_fct<T>(inorm, dims_in, axes);
188 189
  pocketfft::r2c(dims_in, s_in, s_out, axes, forward, d_in, d_out, fct,
    nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
190
  }
Martin Reinecke's avatar
more  
Martin Reinecke committed
191 192
  return res;
  }
193

Martin Reinecke's avatar
Martin Reinecke committed
194 195
py::array r2c(const py::array &in, const py::object &axes_, bool forward,
  int inorm, py::object &out_, size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
196
  {
Martin Reinecke's avatar
Martin Reinecke committed
197
  DISPATCH(in, f64, f32, flong, r2c_internal, (in, axes_, forward, inorm, out_,
Martin Reinecke's avatar
Martin Reinecke committed
198
    nthreads))
Martin Reinecke's avatar
Martin Reinecke committed
199
  }
200

Martin Reinecke's avatar
Martin Reinecke committed
201
template<typename T> py::array r2r_fftpack_internal(const py::array &in,
202
  const py::object &axes_, bool real2hermitian, bool forward, int inorm,
Martin Reinecke's avatar
Martin Reinecke committed
203
  py::object &out_, size_t nthreads)
204
  {
Martin Reinecke's avatar
Martin Reinecke committed
205
  auto axes = makeaxes(in, axes_);
206
  auto dims(copy_shape(in));
Martin Reinecke's avatar
Martin Reinecke committed
207
  py::array res = prepare_output<T>(out_, dims);
Martin Reinecke's avatar
Martin Reinecke committed
208 209 210 211 212 213
  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;
Martin Reinecke's avatar
Martin Reinecke committed
214
  T fct = norm_fct<T>(inorm, dims, axes);
215
  pocketfft::r2r_fftpack(dims, s_in, s_out, axes, real2hermitian, forward,
216
    d_in, d_out, fct, nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
217
  }
218 219
  return res;
  }
220

Martin Reinecke's avatar
Martin Reinecke committed
221
py::array r2r_fftpack(const py::array &in, const py::object &axes_,
222
  bool real2hermitian, bool forward, int inorm, py::object &out_,
Martin Reinecke's avatar
Martin Reinecke committed
223
  size_t nthreads)
224
  {
225 226
  DISPATCH(in, f64, f32, flong, r2r_fftpack_internal, (in, axes_,
    real2hermitian, forward, inorm, out_, nthreads))
227
  }
228

Martin Reinecke's avatar
Martin Reinecke committed
229 230
template<typename T> py::array r2r_dct_internal(const py::array &in,
  const py::object &axes_, int type, int inorm, py::object &out_,
231 232 233 234 235 236 237 238 239 240 241
  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;
242 243 244 245
  T fct = (type==1) ? norm_fct<T>(inorm, dims, axes, 2, -1)
                    : norm_fct<T>(inorm, dims, axes, 2);
  pocketfft::r2r_dct(dims, s_in, s_out, axes, type, d_in, d_out, fct,
    nthreads);
246 247 248 249
  }
  return res;
  }

Martin Reinecke's avatar
Martin Reinecke committed
250 251
py::array r2r_dct(const py::array &in, int type, const py::object &axes_,
  int inorm, py::object &out_, size_t nthreads)
252
  {
Martin Reinecke's avatar
Martin Reinecke committed
253 254 255
  if ((type<1) || (type>4)) throw runtime_error("invalid DCT type");
  DISPATCH(in, f64, f32, flong, r2r_dct_internal, (in, axes_, type, inorm,
    out_, nthreads))
256 257
  }

Martin Reinecke's avatar
Martin Reinecke committed
258 259
template<typename T> py::array r2r_dst_internal(const py::array &in,
  const py::object &axes_, int type, int inorm, py::object &out_,
260 261 262 263 264 265 266 267 268 269 270
  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;
271 272 273 274
  T fct = (type==1) ? norm_fct<T>(inorm, dims, axes, 2, 1)
                    : norm_fct<T>(inorm, dims, axes, 2);
  pocketfft::r2r_dst(dims, s_in, s_out, axes, type, d_in, d_out, fct,
    nthreads);
275 276 277 278
  }
  return res;
  }

Martin Reinecke's avatar
Martin Reinecke committed
279 280
py::array r2r_dst(const py::array &in, int type, const py::object &axes_,
  int inorm, py::object &out_, size_t nthreads)
281
  {
Martin Reinecke's avatar
Martin Reinecke committed
282 283 284
  if ((type<1) || (type>4)) throw runtime_error("invalid DST type");
  DISPATCH(in, f64, f32, flong, r2r_dst_internal, (in, axes_, type, inorm,
    out_, nthreads))
285 286
  }

Martin Reinecke's avatar
Martin Reinecke committed
287
template<typename T> py::array c2r_internal(const py::array &in,
Martin Reinecke's avatar
Martin Reinecke committed
288 289
  const py::object &axes_, size_t lastsize, bool forward, int inorm,
  py::object &out_, size_t nthreads)
Martin Reinecke's avatar
more  
Martin Reinecke committed
290
  {
Martin Reinecke's avatar
Martin Reinecke committed
291
  auto axes = makeaxes(in, axes_);
Martin Reinecke's avatar
more  
Martin Reinecke committed
292
  size_t axis = axes.back();
Martin Reinecke's avatar
Martin Reinecke committed
293 294 295
  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])
Martin Reinecke's avatar
more  
Martin Reinecke committed
296 297
    throw runtime_error("bad lastsize");
  dims_out[axis] = lastsize;
Martin Reinecke's avatar
Martin Reinecke committed
298
  py::array res = prepare_output<T>(out_, dims_out);
Martin Reinecke's avatar
Martin Reinecke committed
299 300 301 302 303 304
  auto s_in=copy_strides(in);
  auto s_out=copy_strides(res);
  auto d_in=reinterpret_cast<const complex<T> *>(in.data());
  auto d_out=reinterpret_cast<T *>(res.mutable_data());
  {
  py::gil_scoped_release release;
305
  T fct = norm_fct<T>(inorm, dims_out, axes);
306 307
  pocketfft::c2r(dims_out, s_in, s_out, axes, forward, d_in, d_out, fct,
    nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
308
  }
Martin Reinecke's avatar
more  
Martin Reinecke committed
309 310
  return res;
  }
311

Martin Reinecke's avatar
Martin Reinecke committed
312 313
py::array c2r(const py::array &in, const py::object &axes_, size_t lastsize,
  bool forward, int inorm, py::object &out_, size_t nthreads)
Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
314
  {
Martin Reinecke's avatar
Martin Reinecke committed
315
  DISPATCH(in, c128, c64, clong, c2r_internal, (in, axes_, lastsize, forward,
Martin Reinecke's avatar
Martin Reinecke committed
316
    inorm, out_, nthreads))
Martin Reinecke's avatar
cleanup  
Martin Reinecke committed
317
  }
Martin Reinecke's avatar
Martin Reinecke committed
318

319
template<typename T> py::array separable_hartley_internal(const py::array &in,
Martin Reinecke's avatar
Martin Reinecke committed
320
  const py::object &axes_, int inorm, py::object &out_, size_t nthreads)
Martin Reinecke's avatar
more  
Martin Reinecke committed
321
  {
Martin Reinecke's avatar
Martin Reinecke committed
322
  auto dims(copy_shape(in));
Martin Reinecke's avatar
Martin Reinecke committed
323
  py::array res = prepare_output<T>(out_, dims);
Martin Reinecke's avatar
Martin Reinecke committed
324 325 326 327 328 329 330
  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;
331
  T fct = norm_fct<T>(inorm, dims, axes);
332 333
  pocketfft::r2r_separable_hartley(dims, s_in, s_out, axes, d_in, d_out, fct,
    nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
334
  }
Martin Reinecke's avatar
Martin Reinecke committed
335 336
  return res;
  }
337

Martin Reinecke's avatar
Martin Reinecke committed
338 339
py::array separable_hartley(const py::array &in, const py::object &axes_,
  int inorm, py::object &out_, size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
340
  {
341
  DISPATCH(in, f64, f32, flong, separable_hartley_internal, (in, axes_, inorm,
Martin Reinecke's avatar
Martin Reinecke committed
342
    out_, nthreads))
Martin Reinecke's avatar
Martin Reinecke committed
343
  }
344

Martin Reinecke's avatar
fixes  
Martin Reinecke committed
345
template<typename T>py::array complex2hartley(const py::array &in,
Martin Reinecke's avatar
Martin Reinecke committed
346
  const py::array &tmp, const py::object &axes_, py::object &out_)
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
347
  {
Martin Reinecke's avatar
Martin Reinecke committed
348
  using namespace pocketfft::detail;
Martin Reinecke's avatar
Martin Reinecke committed
349
  auto dims_out(copy_shape(in));
Martin Reinecke's avatar
Martin Reinecke committed
350
  py::array out = prepare_output<T>(out_, dims_out);
351
  cndarr<cmplx<T>> atmp(tmp.data(), copy_shape(tmp), copy_strides(tmp));
Martin Reinecke's avatar
Martin Reinecke committed
352
  ndarr<T> aout(out.mutable_data(), copy_shape(out), copy_strides(out));
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
353
  auto axes = makeaxes(in, axes_);
Martin Reinecke's avatar
Martin Reinecke committed
354 355
  {
  py::gil_scoped_release release;
356 357
  simple_iter iin(atmp);
  rev_iter iout(aout, axes);
358 359
  if (iin.remaining()!=iout.remaining()) throw runtime_error("oops");
  while(iin.remaining()>0)
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
360
    {
361 362 363
    auto v = atmp[iin.ofs()];
    aout[iout.ofs()] = v.r+v.i;
    aout[iout.rev_ofs()] = v.r-v.i;
364
    iin.advance(); iout.advance();
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
365
    }
Martin Reinecke's avatar
Martin Reinecke committed
366
  }
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
367 368
  return out;
  }
369

Martin Reinecke's avatar
Martin Reinecke committed
370 371
py::array genuine_hartley(const py::array &in, const py::object &axes_,
  int inorm, py::object &out_, size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
372
  {
Martin Reinecke's avatar
Martin Reinecke committed
373 374
  auto tmp = r2c(in, axes_, true, inorm, None, nthreads);
  DISPATCH(in, f64, f32, flong, complex2hartley, (in, tmp, axes_, out_))
Martin Reinecke's avatar
Martin Reinecke committed
375
  }
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
376

Martin Reinecke's avatar
Martin Reinecke committed
377
const char *pypocketfft_DS = R"""(Fast Fourier and Hartley transforms.
Martin Reinecke's avatar
Martin Reinecke committed
378 379

This module supports
380
- single, double, and long double precision
Martin Reinecke's avatar
Martin Reinecke committed
381 382 383 384 385 386
- 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.
Martin Reinecke's avatar
Martin Reinecke committed
387
)""";
Martin Reinecke's avatar
Martin Reinecke committed
388

Martin Reinecke's avatar
Martin Reinecke committed
389
const char *c2c_DS = R"""(Performs a complex FFT.
390 391 392

Parameters
----------
393 394 395
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.
396
axes : list of integers
397
    The axes along which the FFT is carried out.
398
    If not set, all axes will be transformed.
Martin Reinecke's avatar
Martin Reinecke committed
399 400
forward : bool
    If `True`, a negative sign is used in the exponent, else a positive one.
401 402 403 404 405 406
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.
Martin Reinecke's avatar
Martin Reinecke committed
407
out : numpy.ndarray (same shape as `a`, complex type with same accuracy as `a`)
Martin Reinecke's avatar
Martin Reinecke committed
408
    May be identical to `a`, but if it isn't, it must not overlap with `a`.
Martin Reinecke's avatar
Martin Reinecke committed
409
    If None, a new array is allocated to store the output.
Martin Reinecke's avatar
Martin Reinecke committed
410 411 412
nthreads : int
    Number of threads to use. If 0, use the system default (typically governed
    by the `OMP_NUM_THREADS` environment variable).
413 414 415

Returns
-------
Martin Reinecke's avatar
Martin Reinecke committed
416
numpy.ndarray (same shape as `a`, complex type with same accuracy as `a`)
417
    The transformed data.
Martin Reinecke's avatar
Martin Reinecke committed
418
)""";
419

Martin Reinecke's avatar
Martin Reinecke committed
420
const char *r2c_DS = R"""(Performs an FFT whose input is strictly real.
Martin Reinecke's avatar
Martin Reinecke committed
421 422 423

Parameters
----------
424
a : numpy.ndarray (any real type)
Martin Reinecke's avatar
Martin Reinecke committed
425 426
    The input data
axes : list of integers
427
    The axes along which the FFT is carried out.
Martin Reinecke's avatar
Martin Reinecke committed
428
    If not set, all axes will be transformed in ascending order.
Martin Reinecke's avatar
Martin Reinecke committed
429 430
forward : bool
    If `True`, a negative sign is used in the exponent, else a positive one.
431 432 433 434 435 436
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.
Martin Reinecke's avatar
Martin Reinecke committed
437 438
out : numpy.ndarray (complex type with same accuracy as `a`)
    For the required shape, see the `Returns` section.
Martin Reinecke's avatar
Martin Reinecke committed
439
    Must not overlap with `a`.
Martin Reinecke's avatar
Martin Reinecke committed
440
    If None, a new array is allocated to store the output.
Martin Reinecke's avatar
Martin Reinecke committed
441 442 443
nthreads : int
    Number of threads to use. If 0, use the system default (typically governed
    by the `OMP_NUM_THREADS` environment variable).
Martin Reinecke's avatar
Martin Reinecke committed
444 445 446

Returns
-------
Martin Reinecke's avatar
Martin Reinecke committed
447
numpy.ndarray (complex type with same accuracy as `a`)
Martin Reinecke's avatar
Martin Reinecke committed
448
    The transformed data. The shape is identical to that of the input array,
Martin Reinecke's avatar
Martin Reinecke committed
449 450
    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.
Martin Reinecke's avatar
Martin Reinecke committed
451
)""";
Martin Reinecke's avatar
Martin Reinecke committed
452

Martin Reinecke's avatar
Martin Reinecke committed
453
const char *c2r_DS = R"""(Performs an FFT whose output is strictly real.
Martin Reinecke's avatar
Martin Reinecke committed
454 455 456

Parameters
----------
457
a : numpy.ndarray (any complex type)
Martin Reinecke's avatar
Martin Reinecke committed
458 459
    The input data
axes : list of integers
460
    The axes along which the FFT is carried out.
Martin Reinecke's avatar
Martin Reinecke committed
461 462 463
    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.
Martin Reinecke's avatar
Martin Reinecke committed
464 465
forward : bool
    If `True`, a negative sign is used in the exponent, else a positive one.
466 467 468 469 470 471
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.
Martin Reinecke's avatar
Martin Reinecke committed
472 473
out : numpy.ndarray (real type with same accuracy as `a`)
    For the required shape, see the `Returns` section.
Martin Reinecke's avatar
Martin Reinecke committed
474
    Must not overlap with `a`.
Martin Reinecke's avatar
Martin Reinecke committed
475
    If None, a new array is allocated to store the output.
Martin Reinecke's avatar
Martin Reinecke committed
476 477 478
nthreads : int
    Number of threads to use. If 0, use the system default (typically governed
    by the `OMP_NUM_THREADS` environment variable).
Martin Reinecke's avatar
Martin Reinecke committed
479 480 481

Returns
-------
Martin Reinecke's avatar
Martin Reinecke committed
482
numpy.ndarray (real type with same accuracy as `a`)
Martin Reinecke's avatar
Martin Reinecke committed
483 484
    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`
Martin Reinecke's avatar
Martin Reinecke committed
485
    entries.
Martin Reinecke's avatar
Martin Reinecke committed
486
)""";
Martin Reinecke's avatar
Martin Reinecke committed
487

Martin Reinecke's avatar
Martin Reinecke committed
488
const char *r2r_fftpack_DS = R"""(Performs a real-valued FFT using the FFTPACK storage scheme.
489 490 491

Parameters
----------
492
a : numpy.ndarray (any real type)
Martin Reinecke's avatar
Martin Reinecke committed
493 494 495 496
    The input data
axes : list of integers
    The axes along which the FFT is carried out.
    If not set, all axes will be transformed.
497 498 499 500
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.
Martin Reinecke's avatar
Martin Reinecke committed
501 502
forward : bool
    If `True`, a negative sign is used in the exponent, else a positive one.
503 504 505 506 507 508
inorm : int
    Normalization type
      0 : no normalization
      1 : divide by sqrt(N)
      2 : divide by N
    where N is the length of `axis`.
Martin Reinecke's avatar
Martin Reinecke committed
509
out : numpy.ndarray (same shape and data type as `a`)
Martin Reinecke's avatar
Martin Reinecke committed
510
    May be identical to `a`, but if it isn't, it must not overlap with `a`.
Martin Reinecke's avatar
Martin Reinecke committed
511
    If None, a new array is allocated to store the output.
Martin Reinecke's avatar
Martin Reinecke committed
512 513 514
nthreads : int
    Number of threads to use. If 0, use the system default (typically governed
    by the `OMP_NUM_THREADS` environment variable).
515 516 517

Returns
-------
Martin Reinecke's avatar
Martin Reinecke committed
518
numpy.ndarray (same shape and data type as `a`)
519
    The transformed data. The shape is identical to that of the input array.
Martin Reinecke's avatar
Martin Reinecke committed
520
)""";
521

522 523 524 525 526 527 528
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
----------
Martin Reinecke's avatar
Martin Reinecke committed
529
a : numpy.ndarray (any real type)
530 531 532 533 534 535 536 537 538 539
    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.
Martin Reinecke's avatar
Martin Reinecke committed
540
out : numpy.ndarray (same shape and data type as `a`)
Martin Reinecke's avatar
Martin Reinecke committed
541
    May be identical to `a`, but if it isn't, it must not overlap with `a`.
Martin Reinecke's avatar
Martin Reinecke committed
542
    If None, a new array is allocated to store the output.
543 544 545 546 547 548
nthreads : int
    Number of threads to use. If 0, use the system default (typically governed
    by the `OMP_NUM_THREADS` environment variable).

Returns
-------
Martin Reinecke's avatar
Martin Reinecke committed
549
numpy.ndarray (same shape and data type as `a`)
550 551 552 553 554 555 556 557
    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.
558 559 560

Parameters
----------
Martin Reinecke's avatar
Martin Reinecke committed
561
a : numpy.ndarray (any real type)
562 563
    The input data
axes : list of integers
564
    The axes along which the transform is carried out.
565
    If not set, all axes will be transformed.
566 567 568 569 570 571
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.
Martin Reinecke's avatar
Martin Reinecke committed
572
out : numpy.ndarray (same shape and data type as `a`)
Martin Reinecke's avatar
Martin Reinecke committed
573
    May be identical to `a`, but if it isn't, it must not overlap with `a`.
Martin Reinecke's avatar
Martin Reinecke committed
574
    If None, a new array is allocated to store the output.
Martin Reinecke's avatar
Martin Reinecke committed
575 576 577
nthreads : int
    Number of threads to use. If 0, use the system default (typically governed
    by the `OMP_NUM_THREADS` environment variable).
578 579 580

Returns
-------
Martin Reinecke's avatar
Martin Reinecke committed
581
numpy.ndarray (same shape and data type as `a`)
Martin Reinecke's avatar
Martin Reinecke committed
582
    The transformed data
Martin Reinecke's avatar
Martin Reinecke committed
583
)""";
584

Martin Reinecke's avatar
Martin Reinecke committed
585 586 587 588
} // unnamed namespace

PYBIND11_MODULE(pypocketfft, m)
  {
Martin Reinecke's avatar
more  
Martin Reinecke committed
589 590
  using namespace pybind11::literals;

Martin Reinecke's avatar
Martin Reinecke committed
591
  m.doc() = pypocketfft_DS;
592
  m.def("c2c", c2c, c2c_DS, "a"_a, "axes"_a=None, "forward"_a=true,
Martin Reinecke's avatar
Martin Reinecke committed
593
    "inorm"_a=0, "out"_a=None, "nthreads"_a=1);
594
  m.def("r2c", r2c, r2c_DS, "a"_a, "axes"_a=None, "forward"_a=true,
Martin Reinecke's avatar
Martin Reinecke committed
595
    "inorm"_a=0, "out"_a=None, "nthreads"_a=1);
596
  m.def("c2r", c2r, c2r_DS, "a"_a, "axes"_a=None, "lastsize"_a=0,
Martin Reinecke's avatar
Martin Reinecke committed
597
    "forward"_a=true, "inorm"_a=0, "out"_a=None, "nthreads"_a=1);
598
  m.def("r2r_fftpack", r2r_fftpack, r2r_fftpack_DS, "a"_a, "axes"_a,
599
    "real2hermitian"_a, "forward"_a, "inorm"_a=0, "out"_a=None, "nthreads"_a=1);
600
  m.def("separable_hartley", separable_hartley, separable_hartley_DS, "a"_a,
Martin Reinecke's avatar
Martin Reinecke committed
601
    "axes"_a=None, "inorm"_a=0, "out"_a=None, "nthreads"_a=1);
602
  m.def("genuine_hartley", genuine_hartley, genuine_hartley_DS, "a"_a,
Martin Reinecke's avatar
Martin Reinecke committed
603
    "axes"_a=None, "inorm"_a=0, "out"_a=None, "nthreads"_a=1);
Martin Reinecke's avatar
Martin Reinecke committed
604 605 606 607
  m.def("r2r_dct", r2r_dct, /*r2r_dct_DS,*/ "a"_a, "type"_a, "axes"_a=None,
    "inorm"_a=0, "out"_a=None, "nthreads"_a=1);
  m.def("r2r_dst", r2r_dst, /*r2r_dst_DS,*/ "a"_a, "type"_a, "axes"_a=None,
    "inorm"_a=0, "out"_a=None, "nthreads"_a=1);
Martin Reinecke's avatar
Martin Reinecke committed
608
  }