fft.c
Go to the documentation of this file.
1 /*
2  * FFT/IFFT transforms
3  * Copyright (c) 2008 Loren Merritt
4  * Copyright (c) 2002 Fabrice Bellard
5  * Partly based on libdjbfft by D. J. Bernstein
6  *
7  * This file is part of Libav.
8  *
9  * Libav is free software; you can redistribute it and/or
10  * modify it under the terms of the GNU Lesser General Public
11  * License as published by the Free Software Foundation; either
12  * version 2.1 of the License, or (at your option) any later version.
13  *
14  * Libav is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17  * Lesser General Public License for more details.
18  *
19  * You should have received a copy of the GNU Lesser General Public
20  * License along with Libav; if not, write to the Free Software
21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22  */
23 
29 #include <stdlib.h>
30 #include <string.h>
31 #include "libavutil/mathematics.h"
32 #include "fft.h"
33 #include "fft-internal.h"
34 
35 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */
36 #if !CONFIG_HARDCODED_TABLES
37 COSTABLE(16);
38 COSTABLE(32);
39 COSTABLE(64);
40 COSTABLE(128);
41 COSTABLE(256);
42 COSTABLE(512);
43 COSTABLE(1024);
44 COSTABLE(2048);
45 COSTABLE(4096);
46 COSTABLE(8192);
47 COSTABLE(16384);
48 COSTABLE(32768);
49 COSTABLE(65536);
50 #endif
51 COSTABLE_CONST FFTSample * const FFT_NAME(ff_cos_tabs)[] = {
52  NULL, NULL, NULL, NULL,
53  FFT_NAME(ff_cos_16),
54  FFT_NAME(ff_cos_32),
55  FFT_NAME(ff_cos_64),
56  FFT_NAME(ff_cos_128),
57  FFT_NAME(ff_cos_256),
58  FFT_NAME(ff_cos_512),
59  FFT_NAME(ff_cos_1024),
60  FFT_NAME(ff_cos_2048),
61  FFT_NAME(ff_cos_4096),
62  FFT_NAME(ff_cos_8192),
63  FFT_NAME(ff_cos_16384),
64  FFT_NAME(ff_cos_32768),
65  FFT_NAME(ff_cos_65536),
66 };
67 
68 static void ff_fft_permute_c(FFTContext *s, FFTComplex *z);
69 static void ff_fft_calc_c(FFTContext *s, FFTComplex *z);
70 
71 static int split_radix_permutation(int i, int n, int inverse)
72 {
73  int m;
74  if(n <= 2) return i&1;
75  m = n >> 1;
76  if(!(i&m)) return split_radix_permutation(i, m, inverse)*2;
77  m >>= 1;
78  if(inverse == !(i&m)) return split_radix_permutation(i, m, inverse)*4 + 1;
79  else return split_radix_permutation(i, m, inverse)*4 - 1;
80 }
81 
83 {
84 #if !CONFIG_HARDCODED_TABLES
85  int i;
86  int m = 1<<index;
87  double freq = 2*M_PI/m;
88  FFTSample *tab = FFT_NAME(ff_cos_tabs)[index];
89  for(i=0; i<=m/4; i++)
90  tab[i] = FIX15(cos(i*freq));
91  for(i=1; i<m/4; i++)
92  tab[m/2-i] = tab[i];
93 #endif
94 }
95 
96 static const int avx_tab[] = {
97  0, 4, 1, 5, 8, 12, 9, 13, 2, 6, 3, 7, 10, 14, 11, 15
98 };
99 
100 static int is_second_half_of_fft32(int i, int n)
101 {
102  if (n <= 32)
103  return i >= 16;
104  else if (i < n/2)
105  return is_second_half_of_fft32(i, n/2);
106  else if (i < 3*n/4)
107  return is_second_half_of_fft32(i - n/2, n/4);
108  else
109  return is_second_half_of_fft32(i - 3*n/4, n/4);
110 }
111 
113 {
114  int i;
115  int n = 1 << s->nbits;
116 
117  for (i = 0; i < n; i += 16) {
118  int k;
119  if (is_second_half_of_fft32(i, n)) {
120  for (k = 0; k < 16; k++)
121  s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] =
122  i + avx_tab[k];
123 
124  } else {
125  for (k = 0; k < 16; k++) {
126  int j = i + k;
127  j = (j & ~7) | ((j >> 1) & 3) | ((j << 2) & 4);
128  s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] = j;
129  }
130  }
131  }
132 }
133 
134 av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
135 {
136  int i, j, n;
137 
138  if (nbits < 2 || nbits > 16)
139  goto fail;
140  s->nbits = nbits;
141  n = 1 << nbits;
142 
143  s->revtab = av_malloc(n * sizeof(uint16_t));
144  if (!s->revtab)
145  goto fail;
146  s->tmp_buf = av_malloc(n * sizeof(FFTComplex));
147  if (!s->tmp_buf)
148  goto fail;
149  s->inverse = inverse;
151 
153  s->fft_calc = ff_fft_calc_c;
154 #if CONFIG_MDCT
158 #endif
159 
160 #if CONFIG_FFT_FLOAT
161  if (ARCH_ARM) ff_fft_init_arm(s);
163  if (HAVE_MMX) ff_fft_init_mmx(s);
164  if (CONFIG_MDCT) s->mdct_calcw = s->mdct_calc;
165 #else
168 #endif
169 
170  for(j=4; j<=nbits; j++) {
172  }
173 
174  if (s->fft_permutation == FF_FFT_PERM_AVX) {
175  fft_perm_avx(s);
176  } else {
177  for(i=0; i<n; i++) {
178  int j = i;
180  j = (j&~3) | ((j>>1)&1) | ((j<<1)&2);
181  s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = j;
182  }
183  }
184 
185  return 0;
186  fail:
187  av_freep(&s->revtab);
188  av_freep(&s->tmp_buf);
189  return -1;
190 }
191 
193 {
194  int j, np;
195  const uint16_t *revtab = s->revtab;
196  np = 1 << s->nbits;
197  /* TODO: handle split-radix permute in a more optimal way, probably in-place */
198  for(j=0;j<np;j++) s->tmp_buf[revtab[j]] = z[j];
199  memcpy(z, s->tmp_buf, np * sizeof(FFTComplex));
200 }
201 
203 {
204  av_freep(&s->revtab);
205  av_freep(&s->tmp_buf);
206 }
207 
208 #define BUTTERFLIES(a0,a1,a2,a3) {\
209  BF(t3, t5, t5, t1);\
210  BF(a2.re, a0.re, a0.re, t5);\
211  BF(a3.im, a1.im, a1.im, t3);\
212  BF(t4, t6, t2, t6);\
213  BF(a3.re, a1.re, a1.re, t4);\
214  BF(a2.im, a0.im, a0.im, t6);\
215 }
216 
217 // force loading all the inputs before storing any.
218 // this is slightly slower for small data, but avoids store->load aliasing
219 // for addresses separated by large powers of 2.
220 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
221  FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
222  BF(t3, t5, t5, t1);\
223  BF(a2.re, a0.re, r0, t5);\
224  BF(a3.im, a1.im, i1, t3);\
225  BF(t4, t6, t2, t6);\
226  BF(a3.re, a1.re, r1, t4);\
227  BF(a2.im, a0.im, i0, t6);\
228 }
229 
230 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
231  CMUL(t1, t2, a2.re, a2.im, wre, -wim);\
232  CMUL(t5, t6, a3.re, a3.im, wre, wim);\
233  BUTTERFLIES(a0,a1,a2,a3)\
234 }
235 
236 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
237  t1 = a2.re;\
238  t2 = a2.im;\
239  t5 = a3.re;\
240  t6 = a3.im;\
241  BUTTERFLIES(a0,a1,a2,a3)\
242 }
243 
244 /* z[0...8n-1], w[1...2n-1] */
245 #define PASS(name)\
246 static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\
247 {\
248  FFTDouble t1, t2, t3, t4, t5, t6;\
249  int o1 = 2*n;\
250  int o2 = 4*n;\
251  int o3 = 6*n;\
252  const FFTSample *wim = wre+o1;\
253  n--;\
254 \
255  TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
256  TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
257  do {\
258  z += 2;\
259  wre += 2;\
260  wim -= 2;\
261  TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
262  TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
263  } while(--n);\
264 }
265 
266 PASS(pass)
267 #undef BUTTERFLIES
268 #define BUTTERFLIES BUTTERFLIES_BIG
269 PASS(pass_big)
270 
271 #define DECL_FFT(n,n2,n4)\
272 static void fft##n(FFTComplex *z)\
273 {\
274  fft##n2(z);\
275  fft##n4(z+n4*2);\
276  fft##n4(z+n4*3);\
277  pass(z,FFT_NAME(ff_cos_##n),n4/2);\
278 }
279 
280 static void fft4(FFTComplex *z)
281 {
282  FFTDouble t1, t2, t3, t4, t5, t6, t7, t8;
283 
284  BF(t3, t1, z[0].re, z[1].re);
285  BF(t8, t6, z[3].re, z[2].re);
286  BF(z[2].re, z[0].re, t1, t6);
287  BF(t4, t2, z[0].im, z[1].im);
288  BF(t7, t5, z[2].im, z[3].im);
289  BF(z[3].im, z[1].im, t4, t8);
290  BF(z[3].re, z[1].re, t3, t7);
291  BF(z[2].im, z[0].im, t2, t5);
292 }
293 
294 static void fft8(FFTComplex *z)
295 {
296  FFTDouble t1, t2, t3, t4, t5, t6;
297 
298  fft4(z);
299 
300  BF(t1, z[5].re, z[4].re, -z[5].re);
301  BF(t2, z[5].im, z[4].im, -z[5].im);
302  BF(t5, z[7].re, z[6].re, -z[7].re);
303  BF(t6, z[7].im, z[6].im, -z[7].im);
304 
305  BUTTERFLIES(z[0],z[2],z[4],z[6]);
306  TRANSFORM(z[1],z[3],z[5],z[7],sqrthalf,sqrthalf);
307 }
308 
309 #if !CONFIG_SMALL
310 static void fft16(FFTComplex *z)
311 {
312  FFTDouble t1, t2, t3, t4, t5, t6;
313  FFTSample cos_16_1 = FFT_NAME(ff_cos_16)[1];
314  FFTSample cos_16_3 = FFT_NAME(ff_cos_16)[3];
315 
316  fft8(z);
317  fft4(z+8);
318  fft4(z+12);
319 
320  TRANSFORM_ZERO(z[0],z[4],z[8],z[12]);
321  TRANSFORM(z[2],z[6],z[10],z[14],sqrthalf,sqrthalf);
322  TRANSFORM(z[1],z[5],z[9],z[13],cos_16_1,cos_16_3);
323  TRANSFORM(z[3],z[7],z[11],z[15],cos_16_3,cos_16_1);
324 }
325 #else
326 DECL_FFT(16,8,4)
327 #endif
328 DECL_FFT(32,16,8)
329 DECL_FFT(64,32,16)
330 DECL_FFT(128,64,32)
331 DECL_FFT(256,128,64)
332 DECL_FFT(512,256,128)
333 #if !CONFIG_SMALL
334 #define pass pass_big
335 #endif
336 DECL_FFT(1024,512,256)
337 DECL_FFT(2048,1024,512)
338 DECL_FFT(4096,2048,1024)
339 DECL_FFT(8192,4096,2048)
340 DECL_FFT(16384,8192,4096)
341 DECL_FFT(32768,16384,8192)
342 DECL_FFT(65536,32768,16384)
343 
344 static void (* const fft_dispatch[])(FFTComplex*) = {
345  fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
346  fft2048, fft4096, fft8192, fft16384, fft32768, fft65536,
347 };
348 
350 {
351  fft_dispatch[s->nbits-2](z);
352 }
353