|
3828
|
1 /* |
|
|
2 |
|
|
3 This file is part of Octave. |
|
|
4 |
|
|
5 Octave is free software; you can redistribute it and/or modify it |
|
|
6 under the terms of the GNU General Public License as published by the |
|
|
7 Free Software Foundation; either version 2, or (at your option) any |
|
|
8 later version. |
|
|
9 |
|
|
10 Octave is distributed in the hope that it will be useful, but WITHOUT |
|
|
11 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
|
|
12 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
|
|
13 for more details. |
|
|
14 |
|
|
15 You should have received a copy of the GNU General Public License |
|
|
16 along with Octave; see the file COPYING. If not, write to the Free |
|
|
17 Software Foundation, 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. |
|
|
18 |
|
|
19 */ |
|
|
20 |
|
|
21 #ifdef HAVE_CONFIG_H |
|
|
22 #include <config.h> |
|
|
23 #endif |
|
|
24 |
|
4773
|
25 #if defined (HAVE_FFTW3) |
|
3828
|
26 |
|
4775
|
27 #include <iostream> |
|
|
28 #include <vector> |
|
|
29 |
|
3828
|
30 #include "oct-fftw.h" |
|
|
31 #include "lo-error.h" |
|
|
32 |
|
|
33 // Helper class to create and cache fftw plans for both 1d and 2d. This |
|
|
34 // implementation uses FFTW_ESTIMATE to create the plans, which in theory |
|
4773
|
35 // is suboptimal, but provides quite reasonable performance. |
|
|
36 |
|
|
37 // Also note that if FFTW_ESTIMATE is not used the planner in FFTW3 |
|
|
38 // destroys the input and output arrays. So with the form of the |
|
|
39 // current code we definitely want FFTW_ESTIMATE!! However, we use |
|
|
40 // any wsidom that is available, either in a FFTW3 system wide file |
|
|
41 // or as supplied by the user. |
|
|
42 |
|
|
43 // XXX FIXME XXX If we can ensure 16 byte alignment in Array<T> (<T> *data) |
|
|
44 // the FFTW3 can use SIMD instructions for further acceleration. |
|
|
45 |
|
|
46 // Note that it is profitable to store the FFTW3 plans, for small ffts |
|
3828
|
47 |
|
|
48 class |
|
|
49 octave_fftw_planner |
|
|
50 { |
|
|
51 public: |
|
|
52 octave_fftw_planner (); |
|
|
53 |
|
4773
|
54 fftw_plan create_plan (int dir, const int rank, const dim_vector dims, |
|
|
55 int howmany, int stride, int dist, |
|
|
56 const Complex *in, Complex *out); |
|
|
57 fftw_plan create_plan (const int rank, const dim_vector dims, |
|
|
58 int howmany, int stride, int dist, |
|
|
59 const double *in, Complex *out); |
|
3828
|
60 |
|
|
61 private: |
|
|
62 int plan_flags; |
|
|
63 |
|
4773
|
64 // Plan for fft and ifft of complex values |
|
3828
|
65 fftw_plan plan[2]; |
|
4773
|
66 int d[2]; // dist |
|
|
67 int s[2]; // stride |
|
|
68 int r[2]; // rank |
|
|
69 int h[2]; // howmany |
|
|
70 dim_vector n[2]; // dims |
|
|
71 char ialign[2]; |
|
|
72 char oalign[2]; |
|
3828
|
73 |
|
4773
|
74 // Plan for fft of real values |
|
|
75 fftw_plan rplan; |
|
|
76 int rd; // dist |
|
|
77 int rs; // stride |
|
|
78 int rr; // rank |
|
|
79 int rh; // howmany |
|
|
80 dim_vector rn; // dims |
|
|
81 char rialign; |
|
|
82 char roalign; |
|
3828
|
83 }; |
|
|
84 |
|
|
85 octave_fftw_planner::octave_fftw_planner () |
|
|
86 { |
|
|
87 plan_flags = FFTW_ESTIMATE; |
|
|
88 |
|
|
89 plan[0] = plan[1] = 0; |
|
4773
|
90 d[0] = d[1] = s[0] = s[1] = r[0] = r[1] = h[0] = h[1] = 0; |
|
|
91 ialign[0] = ialign[1] = oalign[0] = oalign[1] = 0; |
|
|
92 n[0] = n[1] = dim_vector(); |
|
|
93 |
|
|
94 rplan = 0; |
|
|
95 rd = rs = rr = rh = 0; |
|
|
96 rialign = roalign = 0; |
|
|
97 rn = dim_vector (); |
|
|
98 |
|
|
99 // If we have a system wide wisdom file, import it |
|
|
100 fftw_import_system_wisdom ( ); |
|
3828
|
101 } |
|
|
102 |
|
|
103 fftw_plan |
|
4773
|
104 octave_fftw_planner::create_plan (int dir, const int rank, |
|
|
105 const dim_vector dims, int howmany, |
|
|
106 int stride, int dist, |
|
|
107 const Complex *in, Complex *out) |
|
3828
|
108 { |
|
4773
|
109 int which = (dir == FFTW_FORWARD) ? 0 : 1; |
|
3828
|
110 fftw_plan *cur_plan_p = &plan[which]; |
|
|
111 bool create_new_plan = false; |
|
4783
|
112 char in_align = (reinterpret_cast<long> (in)) & 0xF; |
|
|
113 char out_align = (reinterpret_cast<long> (out)) & 0xF; |
|
3828
|
114 |
|
4783
|
115 if (plan[which] == 0 || d[which] != dist || s[which] != stride |
|
|
116 || r[which] != rank || h[which] != howmany |
|
|
117 || ialign[which] != in_align || oalign[which] != out_align) |
|
4773
|
118 create_new_plan = true; |
|
|
119 else |
|
|
120 // We still might not have the same shape of array |
|
|
121 for (int i = 0; i < rank; i++) |
|
|
122 if (dims(i) != n[which](i)) |
|
|
123 { |
|
|
124 create_new_plan = true; |
|
|
125 break; |
|
|
126 } |
|
3828
|
127 |
|
|
128 if (create_new_plan) |
|
|
129 { |
|
4773
|
130 d[which] = dist; |
|
|
131 s[which] = stride; |
|
|
132 r[which] = rank; |
|
|
133 h[which] = howmany; |
|
|
134 ialign[which] = in_align; |
|
|
135 oalign[which] = out_align; |
|
|
136 n[which] = dims; |
|
|
137 |
|
3828
|
138 if (*cur_plan_p) |
|
|
139 fftw_destroy_plan (*cur_plan_p); |
|
|
140 |
|
4773
|
141 // Note reversal of dimensions for column major storage in FFTW |
|
|
142 OCTAVE_LOCAL_BUFFER (int, tmp, rank); |
|
|
143 for (int i = 0, j = rank-1; i < rank; i++, j--) |
|
|
144 tmp[i] = dims(j); |
|
|
145 |
|
|
146 *cur_plan_p = |
|
|
147 fftw_plan_many_dft (rank, tmp, howmany, |
|
|
148 reinterpret_cast<fftw_complex *> (const_cast<Complex *> (in)), |
|
4774
|
149 0, stride, dist, reinterpret_cast<fftw_complex *> (out), |
|
|
150 0, stride, dist, dir, plan_flags); |
|
3828
|
151 |
|
|
152 if (*cur_plan_p == 0) |
|
|
153 (*current_liboctave_error_handler) ("Error creating fftw plan"); |
|
|
154 } |
|
|
155 |
|
|
156 return *cur_plan_p; |
|
|
157 } |
|
|
158 |
|
4773
|
159 fftw_plan |
|
|
160 octave_fftw_planner::create_plan (const int rank, const dim_vector dims, |
|
|
161 int howmany, int stride, int dist, |
|
|
162 const double *in, Complex *out) |
|
3828
|
163 { |
|
4773
|
164 fftw_plan *cur_plan_p = &rplan; |
|
3828
|
165 bool create_new_plan = false; |
|
4783
|
166 char in_align = (reinterpret_cast<long> (in)) & 0xF; |
|
|
167 char out_align = (reinterpret_cast<long> (out)) & 0xF; |
|
3828
|
168 |
|
4783
|
169 if (rplan == 0 || rd != dist || rs != stride || rr != rank |
|
|
170 || rh != howmany || rialign != in_align || roalign != out_align) |
|
4773
|
171 create_new_plan = true; |
|
|
172 else |
|
|
173 // We still might not have the same shape of array |
|
|
174 for (int i = 0; i < rank; i++) |
|
|
175 if (dims(i) != rn(i)) |
|
|
176 { |
|
|
177 create_new_plan = true; |
|
|
178 break; |
|
|
179 } |
|
3828
|
180 |
|
|
181 if (create_new_plan) |
|
|
182 { |
|
4773
|
183 rd = dist; |
|
|
184 rs = stride; |
|
|
185 rr = rank; |
|
|
186 rh = howmany; |
|
|
187 rialign = in_align; |
|
|
188 roalign = out_align; |
|
|
189 rn = dims; |
|
|
190 |
|
3828
|
191 if (*cur_plan_p) |
|
4773
|
192 fftw_destroy_plan (*cur_plan_p); |
|
3828
|
193 |
|
4773
|
194 // Note reversal of dimensions for column major storage in FFTW |
|
|
195 OCTAVE_LOCAL_BUFFER (int, tmp, rank); |
|
|
196 for (int i = 0, j = rank-1; i < rank; i++, j--) |
|
|
197 tmp[i] = dims(j); |
|
|
198 |
|
|
199 *cur_plan_p = |
|
|
200 fftw_plan_many_dft_r2c (rank, tmp, howmany, |
|
|
201 (const_cast<double *> (in)), |
|
4774
|
202 0, stride, dist, reinterpret_cast<fftw_complex *> (out), |
|
|
203 0, stride, dist, plan_flags); |
|
3828
|
204 |
|
|
205 if (*cur_plan_p == 0) |
|
4773
|
206 (*current_liboctave_error_handler) ("Error creating fftw plan"); |
|
3828
|
207 } |
|
|
208 |
|
|
209 return *cur_plan_p; |
|
|
210 } |
|
|
211 |
|
|
212 static octave_fftw_planner fftw_planner; |
|
|
213 |
|
4775
|
214 static inline void |
|
|
215 convert_packcomplex_1d (Complex *out, size_t nr, size_t nc, |
|
|
216 int stride, int dist) |
|
4773
|
217 { |
|
4785
|
218 OCTAVE_QUIT; |
|
|
219 |
|
|
220 // Fill in the missing data. |
|
|
221 |
|
4773
|
222 for (size_t i = 0; i < nr; i++) |
|
|
223 for (size_t j = nc/2+1; j < nc; j++) |
|
|
224 out[j*stride + i*dist] = conj(out[(nc - j)*stride + i*dist]); |
|
4785
|
225 |
|
|
226 OCTAVE_QUIT; |
|
4773
|
227 } |
|
|
228 |
|
4775
|
229 static inline void |
|
|
230 convert_packcomplex_Nd (Complex *out, const dim_vector &dv) |
|
3828
|
231 { |
|
4773
|
232 size_t nc = dv(0); |
|
|
233 size_t nr = dv(1); |
|
|
234 size_t np = (dv.length() > 2 ? dv.numel () / nc / nr : 1); |
|
|
235 size_t nrp = nr * np; |
|
|
236 Complex *ptr1, *ptr2; |
|
|
237 |
|
4785
|
238 OCTAVE_QUIT; |
|
|
239 |
|
|
240 // Create space for the missing elements. |
|
|
241 |
|
4773
|
242 for (size_t i = 0; i < nrp; i++) |
|
|
243 { |
|
|
244 ptr1 = out + i * (nc/2 + 1) + nrp*((nc-1)/2); |
|
|
245 ptr2 = out + i * nc; |
|
|
246 for (size_t j = 0; j < nc/2+1; j++) |
|
|
247 *ptr2++ = *ptr1++; |
|
|
248 } |
|
|
249 |
|
4785
|
250 OCTAVE_QUIT; |
|
|
251 |
|
|
252 // Fill in the missing data for the rank = 2 case directly for speed. |
|
|
253 |
|
4773
|
254 for (size_t i = 0; i < np; i++) |
|
|
255 { |
|
|
256 for (size_t j = 1; j < nr; j++) |
|
|
257 for (size_t k = nc/2+1; k < nc; k++) |
|
|
258 out[k + (j + i*nr)*nc] = conj(out[nc - k + ((i+1)*nr - j)*nc]); |
|
|
259 |
|
|
260 for (size_t j = nc/2+1; j < nc; j++) |
|
|
261 out[j + i*nr*nc] = conj(out[(i*nr+1)*nc - j]); |
|
|
262 } |
|
|
263 |
|
4785
|
264 OCTAVE_QUIT; |
|
|
265 |
|
|
266 // Now do the permutations needed for rank > 2 cases. |
|
|
267 |
|
4773
|
268 size_t jstart = dv(0) * dv(1); |
|
|
269 size_t kstep = dv(0); |
|
|
270 size_t nel = dv.numel (); |
|
4785
|
271 |
|
4773
|
272 for (int inner = 2; inner < dv.length(); inner++) |
|
|
273 { |
|
|
274 size_t jmax = jstart * dv(inner); |
|
|
275 for (size_t i = 0; i < nel; i+=jmax) |
|
|
276 for (size_t j = jstart, jj = jmax-jstart; j < jj; |
|
|
277 j+=jstart, jj-=jstart) |
|
|
278 for (size_t k = 0; k < jstart; k+= kstep) |
|
|
279 for (size_t l = nc/2+1; l < nc; l++) |
|
|
280 { |
|
|
281 Complex tmp = out[i+ j + k + l]; |
|
|
282 out[i + j + k + l] = out[i + jj + k + l]; |
|
|
283 out[i + jj + k + l] = tmp; |
|
|
284 } |
|
|
285 jstart = jmax; |
|
|
286 } |
|
4785
|
287 |
|
|
288 OCTAVE_QUIT; |
|
4773
|
289 } |
|
|
290 |
|
|
291 int |
|
|
292 octave_fftw::fft (const double *in, Complex *out, size_t npts, |
|
|
293 size_t nsamples, int stride, int dist) |
|
|
294 { |
|
|
295 dist = (dist < 0 ? npts : dist); |
|
|
296 |
|
|
297 dim_vector dv (npts); |
|
|
298 fftw_plan plan = fftw_planner.create_plan (1, dv, nsamples, stride, dist, |
|
|
299 in, out); |
|
|
300 |
|
|
301 fftw_execute_dft_r2c (plan, (const_cast<double *>(in)), |
|
|
302 reinterpret_cast<fftw_complex *> (out)); |
|
|
303 |
|
|
304 // Need to create other half of the transform |
|
|
305 convert_packcomplex_1d (out, nsamples, npts, stride, dist); |
|
3828
|
306 |
|
|
307 return 0; |
|
|
308 } |
|
|
309 |
|
|
310 int |
|
4773
|
311 octave_fftw::fft (const Complex *in, Complex *out, size_t npts, |
|
|
312 size_t nsamples, int stride, int dist) |
|
3828
|
313 { |
|
4773
|
314 dist = (dist < 0 ? npts : dist); |
|
|
315 |
|
|
316 dim_vector dv (npts); |
|
|
317 fftw_plan plan = fftw_planner.create_plan (FFTW_FORWARD, 1, dv, nsamples, |
|
|
318 stride, dist, in, out); |
|
|
319 |
|
|
320 fftw_execute_dft (plan, |
|
|
321 reinterpret_cast<fftw_complex *> (const_cast<Complex *>(in)), |
|
|
322 reinterpret_cast<fftw_complex *> (out)); |
|
|
323 |
|
|
324 return 0; |
|
|
325 } |
|
|
326 |
|
|
327 int |
|
|
328 octave_fftw::ifft (const Complex *in, Complex *out, size_t npts, |
|
|
329 size_t nsamples, int stride, int dist) |
|
|
330 { |
|
|
331 dist = (dist < 0 ? npts : dist); |
|
|
332 |
|
|
333 dim_vector dv (npts); |
|
|
334 fftw_plan plan = fftw_planner.create_plan (FFTW_BACKWARD, 1, dv, nsamples, |
|
|
335 stride, dist, in, out); |
|
|
336 |
|
|
337 fftw_execute_dft (plan, |
|
|
338 reinterpret_cast<fftw_complex *> (const_cast<Complex *>(in)), |
|
|
339 reinterpret_cast<fftw_complex *> (out)); |
|
3828
|
340 |
|
|
341 const Complex scale = npts; |
|
4773
|
342 for (size_t j = 0; j < nsamples; j++) |
|
|
343 for (size_t i = 0; i < npts; i++) |
|
|
344 out[i*stride + j*dist] /= scale; |
|
3828
|
345 |
|
|
346 return 0; |
|
|
347 } |
|
|
348 |
|
|
349 int |
|
4773
|
350 octave_fftw::fftNd (const double *in, Complex *out, const int rank, |
|
|
351 const dim_vector &dv) |
|
3828
|
352 { |
|
4773
|
353 int dist = 1; |
|
|
354 for (int i = 0; i < rank; i++) |
|
|
355 dist *= dv(i); |
|
|
356 |
|
|
357 // Fool with the position of the start of the output matrix, so that |
|
|
358 // creating other half of the matrix won't cause cache problems |
|
|
359 int offset = (dv.numel () / dv(0)) * ((dv(0) - 1) / 2); |
|
|
360 |
|
|
361 fftw_plan plan = fftw_planner.create_plan (rank, dv, 1, 1, dist, |
|
|
362 in, out + offset); |
|
|
363 |
|
|
364 fftw_execute_dft_r2c (plan, (const_cast<double *>(in)), |
|
|
365 reinterpret_cast<fftw_complex *> (out+ offset)); |
|
|
366 |
|
|
367 // Need to create other half of the transform |
|
|
368 convert_packcomplex_Nd (out, dv); |
|
3828
|
369 |
|
|
370 return 0; |
|
|
371 } |
|
|
372 |
|
|
373 int |
|
4773
|
374 octave_fftw::fftNd (const Complex *in, Complex *out, const int rank, |
|
|
375 const dim_vector &dv) |
|
3828
|
376 { |
|
4773
|
377 int dist = 1; |
|
|
378 for (int i = 0; i < rank; i++) |
|
|
379 dist *= dv(i); |
|
|
380 |
|
|
381 fftw_plan plan = fftw_planner.create_plan (FFTW_FORWARD, rank, dv, 1, 1, |
|
|
382 dist, in, out); |
|
|
383 |
|
|
384 fftw_execute_dft (plan, |
|
|
385 reinterpret_cast<fftw_complex *> (const_cast<Complex *>(in)), |
|
|
386 reinterpret_cast<fftw_complex *> (out)); |
|
|
387 |
|
|
388 return 0; |
|
|
389 } |
|
3828
|
390 |
|
4773
|
391 int |
|
|
392 octave_fftw::ifftNd (const Complex *in, Complex *out, const int rank, |
|
4784
|
393 const dim_vector &dv) |
|
4773
|
394 { |
|
|
395 int dist = 1; |
|
|
396 for (int i = 0; i < rank; i++) |
|
|
397 dist *= dv(i); |
|
|
398 |
|
|
399 fftw_plan plan = fftw_planner.create_plan (FFTW_BACKWARD, rank, dv, 1, 1, |
|
|
400 dist, in, out); |
|
|
401 |
|
|
402 fftw_execute_dft (plan, |
|
|
403 reinterpret_cast<fftw_complex *> (const_cast<Complex *>(in)), |
|
|
404 reinterpret_cast<fftw_complex *> (out)); |
|
|
405 |
|
|
406 const size_t npts = dv.numel (); |
|
3828
|
407 const Complex scale = npts; |
|
|
408 for (size_t i = 0; i < npts; i++) |
|
4773
|
409 out[i] /= scale; |
|
3828
|
410 |
|
|
411 return 0; |
|
|
412 } |
|
|
413 |
|
|
414 #endif |
|
|
415 |
|
|
416 /* |
|
|
417 ;;; Local Variables: *** |
|
|
418 ;;; mode: C++ *** |
|
|
419 ;;; End: *** |
|
|
420 */ |
|
|
421 |
