Rev 2 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
2 | pj | 1 | /* |
2 | * Copyright (c) 1997-1999 Massachusetts Institute of Technology |
||
3 | * |
||
4 | * This program is free software; you can redistribute it and/or modify |
||
5 | * it under the terms of the GNU General Public License as published by |
||
6 | * the Free Software Foundation; either version 2 of the License, or |
||
7 | * (at your option) any later version. |
||
8 | * |
||
9 | * This program is distributed in the hope that it will be useful, |
||
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
||
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
||
12 | * GNU General Public License for more details. |
||
13 | * |
||
14 | * You should have received a copy of the GNU General Public License |
||
15 | * along with this program; if not, write to the Free Software |
||
16 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
||
17 | * |
||
18 | */ |
||
19 | |||
20 | /* |
||
21 | * Compute transforms of prime sizes using Rader's trick: turn them |
||
22 | * into convolutions of size n - 1, which you then perform via a pair |
||
23 | * of FFTs. |
||
24 | */ |
||
25 | |||
26 | #include <stdlib.h> |
||
27 | #include <math.h> |
||
28 | |||
29 | #include <ports/fftw-int.h> |
||
30 | |||
31 | #ifdef FFTW_USING_CILK |
||
32 | #include <ports/cilk.h> |
||
33 | #include <ports/cilk-compat.h> |
||
34 | #endif |
||
35 | |||
36 | #ifdef FFTW_DEBUG |
||
37 | #define WHEN_DEBUG(a) a |
||
38 | #else |
||
39 | #define WHEN_DEBUG(a) |
||
40 | #endif |
||
41 | |||
42 | /* compute n^m mod p, where m >= 0 and p > 0. */ |
||
43 | static int power_mod(int n, int m, int p) |
||
44 | { |
||
45 | if (m == 0) |
||
46 | return 1; |
||
47 | else if (m % 2 == 0) { |
||
48 | int x = power_mod(n, m / 2, p); |
||
49 | return ((x * x) % p); |
||
50 | } else |
||
51 | return ((n * power_mod(n, m - 1, p)) % p); |
||
52 | } |
||
53 | |||
54 | /* |
||
55 | * Find the period of n in the multiplicative group mod p (p prime). |
||
56 | * That is, return the smallest m such that n^m == 1 mod p. |
||
57 | */ |
||
58 | static int period(int n, int p) |
||
59 | { |
||
60 | int prod = n, period = 1; |
||
61 | |||
62 | while (prod != 1) { |
||
63 | prod = (prod * n) % p; |
||
64 | ++period; |
||
65 | if (prod == 0) |
||
66 | fftw_die("non-prime order in Rader\n"); |
||
67 | } |
||
68 | return period; |
||
69 | } |
||
70 | |||
71 | /* find a generator for the multiplicative group mod p, where p is prime */ |
||
72 | static int find_generator(int p) |
||
73 | { |
||
74 | int g; |
||
75 | |||
76 | for (g = 1; g < p; ++g) |
||
77 | if (period(g, p) == p - 1) |
||
78 | break; |
||
79 | if (g == p) |
||
80 | fftw_die("couldn't find generator for Rader\n"); |
||
81 | return g; |
||
82 | } |
||
83 | |||
84 | /***************************************************************************/ |
||
85 | |||
86 | static fftw_rader_data *create_rader_aux(int p, int flags) |
||
87 | { |
||
88 | fftw_complex *omega, *work; |
||
89 | int g, ginv, gpower; |
||
90 | int i; |
||
91 | FFTW_TRIG_REAL twoPiOverN; |
||
92 | fftw_real scale = 1.0 / (p - 1); /* for convolution */ |
||
93 | fftw_plan plan; |
||
94 | fftw_rader_data *d; |
||
95 | |||
96 | if (p < 2) |
||
97 | fftw_die("non-prime order in Rader\n"); |
||
98 | |||
99 | flags &= ~FFTW_IN_PLACE; |
||
100 | |||
101 | d = (fftw_rader_data *) fftw_malloc(sizeof(fftw_rader_data)); |
||
102 | |||
103 | g = find_generator(p); |
||
104 | ginv = power_mod(g, p - 2, p); |
||
105 | |||
106 | omega = (fftw_complex *) fftw_malloc((p - 1) * sizeof(fftw_complex)); |
||
107 | |||
108 | plan = fftw_create_plan(p - 1, FFTW_FORWARD, flags); |
||
109 | |||
110 | work = (fftw_complex *) fftw_malloc((p - 1) * sizeof(fftw_complex)); |
||
111 | |||
112 | twoPiOverN = FFTW_K2PI / (FFTW_TRIG_REAL) p; |
||
113 | gpower = 1; |
||
114 | for (i = 0; i < p - 1; ++i) { |
||
115 | c_re(work[i]) = scale * FFTW_TRIG_COS(twoPiOverN * gpower); |
||
116 | c_im(work[i]) = FFTW_FORWARD * scale * FFTW_TRIG_SIN(twoPiOverN * gpower); |
||
117 | gpower = (gpower * ginv) % p; |
||
118 | } |
||
119 | |||
120 | /* fft permuted roots of unity */ |
||
121 | fftw_executor_simple(p - 1, work, omega, plan->root, 1, 1); |
||
122 | |||
123 | fftw_free(work); |
||
124 | |||
125 | d->plan = plan; |
||
126 | d->omega = omega; |
||
127 | d->g = g; |
||
128 | d->ginv = ginv; |
||
129 | d->p = p; |
||
130 | d->flags = flags; |
||
131 | d->refcount = 1; |
||
132 | d->next = NULL; |
||
133 | |||
134 | d->cdesc = (fftw_codelet_desc *) fftw_malloc(sizeof(fftw_codelet_desc)); |
||
135 | d->cdesc->name = NULL; |
||
136 | d->cdesc->codelet = NULL; |
||
137 | d->cdesc->size = p; |
||
138 | d->cdesc->dir = FFTW_FORWARD; |
||
139 | d->cdesc->type = FFTW_RADER; |
||
140 | d->cdesc->signature = g; |
||
141 | d->cdesc->ntwiddle = 0; |
||
142 | d->cdesc->twiddle_order = NULL; |
||
143 | return d; |
||
144 | } |
||
145 | |||
146 | /***************************************************************************/ |
||
147 | |||
148 | static fftw_rader_data *fftw_create_rader(int p, int flags) |
||
149 | { |
||
150 | fftw_rader_data *d = fftw_rader_top; |
||
151 | |||
152 | flags &= ~FFTW_IN_PLACE; |
||
153 | while (d && (d->p != p || d->flags != flags)) |
||
154 | d = d->next; |
||
155 | if (d) { |
||
156 | d->refcount++; |
||
157 | return d; |
||
158 | } |
||
159 | d = create_rader_aux(p, flags); |
||
160 | d->next = fftw_rader_top; |
||
161 | fftw_rader_top = d; |
||
162 | return d; |
||
163 | } |
||
164 | |||
165 | /***************************************************************************/ |
||
166 | |||
167 | /* Compute the prime FFTs, premultiplied by twiddle factors. Below, we |
||
168 | * extensively use the identity that fft(x*)* = ifft(x) in order to |
||
169 | * share data between forward and backward transforms and to obviate |
||
170 | * the necessity of having separate forward and backward plans. */ |
||
171 | |||
172 | void fftw_twiddle_rader(fftw_complex *A, const fftw_complex *W, |
||
173 | int m, int r, int stride, |
||
174 | fftw_rader_data * d) |
||
175 | { |
||
176 | fftw_complex *tmp = (fftw_complex *) |
||
177 | fftw_malloc((r - 1) * sizeof(fftw_complex)); |
||
178 | int i, k, gpower = 1, g = d->g, ginv = d->ginv; |
||
179 | fftw_real a0r, a0i; |
||
180 | fftw_complex *omega = d->omega; |
||
181 | |||
182 | for (i = 0; i < m; ++i, A += stride, W += r - 1) { |
||
183 | /* |
||
184 | * Here, we fft W[k-1] * A[k*(m*stride)], using Rader. |
||
185 | * (Actually, W is pre-permuted to match the permutation that we |
||
186 | * will do on A.) |
||
187 | */ |
||
188 | |||
189 | /* First, permute the input and multiply by W, storing in tmp: */ |
||
190 | /* gpower == g^k mod r in the following loop */ |
||
191 | for (k = 0; k < r - 1; ++k, gpower = (gpower * g) % r) { |
||
192 | fftw_real rA, iA, rW, iW; |
||
193 | rW = c_re(W[k]); |
||
194 | iW = c_im(W[k]); |
||
195 | rA = c_re(A[gpower * (m * stride)]); |
||
196 | iA = c_im(A[gpower * (m * stride)]); |
||
197 | c_re(tmp[k]) = rW * rA - iW * iA; |
||
198 | c_im(tmp[k]) = rW * iA + iW * rA; |
||
199 | } |
||
200 | |||
201 | WHEN_DEBUG( { |
||
202 | if (gpower != 1) |
||
203 | fftw_die("incorrect generator in Rader\n"); |
||
204 | } |
||
205 | ); |
||
206 | |||
207 | /* FFT tmp to A: */ |
||
208 | fftw_executor_simple(r - 1, tmp, A + (m * stride), |
||
209 | d->plan->root, 1, m * stride); |
||
210 | |||
211 | /* set output DC component: */ |
||
212 | a0r = c_re(A[0]); |
||
213 | a0i = c_im(A[0]); |
||
214 | c_re(A[0]) += c_re(A[(m * stride)]); |
||
215 | c_im(A[0]) += c_im(A[(m * stride)]); |
||
216 | |||
217 | /* now, multiply by omega: */ |
||
218 | for (k = 0; k < r - 1; ++k) { |
||
219 | fftw_real rA, iA, rW, iW; |
||
220 | rW = c_re(omega[k]); |
||
221 | iW = c_im(omega[k]); |
||
222 | rA = c_re(A[(k + 1) * (m * stride)]); |
||
223 | iA = c_im(A[(k + 1) * (m * stride)]); |
||
224 | c_re(A[(k + 1) * (m * stride)]) = rW * rA - iW * iA; |
||
225 | c_im(A[(k + 1) * (m * stride)]) = -(rW * iA + iW * rA); |
||
226 | } |
||
227 | |||
228 | /* this will add A[0] to all of the outputs after the ifft */ |
||
229 | c_re(A[(m * stride)]) += a0r; |
||
230 | c_im(A[(m * stride)]) -= a0i; |
||
231 | |||
232 | /* inverse FFT: */ |
||
233 | fftw_executor_simple(r - 1, A + (m * stride), tmp, |
||
234 | d->plan->root, m * stride, 1); |
||
235 | |||
236 | /* finally, do inverse permutation to unshuffle the output: */ |
||
237 | for (k = 0; k < r - 1; ++k, gpower = (gpower * ginv) % r) { |
||
238 | c_re(A[gpower * (m * stride)]) = c_re(tmp[k]); |
||
239 | c_im(A[gpower * (m * stride)]) = -c_im(tmp[k]); |
||
240 | } |
||
241 | |||
242 | WHEN_DEBUG( { |
||
243 | if (gpower != 1) |
||
244 | fftw_die("incorrect generator in Rader\n"); |
||
245 | } |
||
246 | ); |
||
247 | |||
248 | } |
||
249 | |||
250 | fftw_free(tmp); |
||
251 | } |
||
252 | |||
253 | void fftwi_twiddle_rader(fftw_complex *A, const fftw_complex *W, |
||
254 | int m, int r, int stride, |
||
255 | fftw_rader_data * d) |
||
256 | { |
||
257 | fftw_complex *tmp = (fftw_complex *) |
||
258 | fftw_malloc((r - 1) * sizeof(fftw_complex)); |
||
259 | int i, k, gpower = 1, g = d->g, ginv = d->ginv; |
||
260 | fftw_real a0r, a0i; |
||
261 | fftw_complex *omega = d->omega; |
||
262 | |||
263 | for (i = 0; i < m; ++i, A += stride, W += r - 1) { |
||
264 | /* |
||
265 | * Here, we fft W[k-1]* * A[k*(m*stride)], using Rader. |
||
266 | * (Actually, W is pre-permuted to match the permutation that |
||
267 | * we will do on A.) |
||
268 | */ |
||
269 | |||
270 | /* First, permute the input and multiply by W*, storing in tmp: */ |
||
271 | /* gpower == g^k mod r in the following loop */ |
||
272 | for (k = 0; k < r - 1; ++k, gpower = (gpower * g) % r) { |
||
273 | fftw_real rA, iA, rW, iW; |
||
274 | rW = c_re(W[k]); |
||
275 | iW = c_im(W[k]); |
||
276 | rA = c_re(A[gpower * (m * stride)]); |
||
277 | iA = c_im(A[gpower * (m * stride)]); |
||
278 | c_re(tmp[k]) = rW * rA + iW * iA; |
||
279 | c_im(tmp[k]) = iW * rA - rW * iA; |
||
280 | } |
||
281 | |||
282 | WHEN_DEBUG( { |
||
283 | if (gpower != 1) |
||
284 | fftw_die("incorrect generator in Rader\n"); |
||
285 | } |
||
286 | ); |
||
287 | |||
288 | /* FFT tmp to A: */ |
||
289 | fftw_executor_simple(r - 1, tmp, A + (m * stride), |
||
290 | d->plan->root, 1, m * stride); |
||
291 | |||
292 | /* set output DC component: */ |
||
293 | a0r = c_re(A[0]); |
||
294 | a0i = c_im(A[0]); |
||
295 | c_re(A[0]) += c_re(A[(m * stride)]); |
||
296 | c_im(A[0]) -= c_im(A[(m * stride)]); |
||
297 | |||
298 | /* now, multiply by omega: */ |
||
299 | for (k = 0; k < r - 1; ++k) { |
||
300 | fftw_real rA, iA, rW, iW; |
||
301 | rW = c_re(omega[k]); |
||
302 | iW = c_im(omega[k]); |
||
303 | rA = c_re(A[(k + 1) * (m * stride)]); |
||
304 | iA = c_im(A[(k + 1) * (m * stride)]); |
||
305 | c_re(A[(k + 1) * (m * stride)]) = rW * rA - iW * iA; |
||
306 | c_im(A[(k + 1) * (m * stride)]) = -(rW * iA + iW * rA); |
||
307 | } |
||
308 | |||
309 | /* this will add A[0] to all of the outputs after the ifft */ |
||
310 | c_re(A[(m * stride)]) += a0r; |
||
311 | c_im(A[(m * stride)]) += a0i; |
||
312 | |||
313 | /* inverse FFT: */ |
||
314 | fftw_executor_simple(r - 1, A + (m * stride), tmp, |
||
315 | d->plan->root, m * stride, 1); |
||
316 | |||
317 | /* finally, do inverse permutation to unshuffle the output: */ |
||
318 | for (k = 0; k < r - 1; ++k, gpower = (gpower * ginv) % r) { |
||
319 | A[gpower * (m * stride)] = tmp[k]; |
||
320 | } |
||
321 | |||
322 | WHEN_DEBUG( { |
||
323 | if (gpower != 1) |
||
324 | fftw_die("incorrect generator in Rader\n"); |
||
325 | } |
||
326 | ); |
||
327 | } |
||
328 | |||
329 | fftw_free(tmp); |
||
330 | } |
||
331 | |||
332 | /***************************************************************************/ |
||
333 | |||
334 | /* |
||
335 | * Make an FFTW_RADER plan node. Note that this function must go |
||
336 | * here, rather than in putils.c, because it indirectly calls the |
||
337 | * fftw_planner. If we included it in putils.c, which is also used |
||
338 | * by rfftw, then any program using rfftw would be linked with all |
||
339 | * of the FFTW codelets, even if they were not needed. I wish that the |
||
340 | * darn linkers operated on a function rather than a file granularity. |
||
341 | */ |
||
342 | fftw_plan_node *fftw_make_node_rader(int n, int size, fftw_direction dir, |
||
343 | fftw_plan_node *recurse, |
||
344 | int flags) |
||
345 | { |
||
346 | fftw_plan_node *p = fftw_make_node(); |
||
347 | |||
348 | p->type = FFTW_RADER; |
||
349 | p->nodeu.rader.size = size; |
||
350 | p->nodeu.rader.codelet = dir == FFTW_FORWARD ? |
||
351 | fftw_twiddle_rader : fftwi_twiddle_rader; |
||
352 | p->nodeu.rader.rader_data = fftw_create_rader(size, flags); |
||
353 | p->nodeu.rader.recurse = recurse; |
||
354 | fftw_use_node(recurse); |
||
355 | |||
356 | if (flags & FFTW_MEASURE) |
||
357 | p->nodeu.rader.tw = |
||
358 | fftw_create_twiddle(n, p->nodeu.rader.rader_data->cdesc); |
||
359 | else |
||
360 | p->nodeu.rader.tw = 0; |
||
361 | return p; |
||
362 | } |