Mercurial > libavcodec.hg
comparison fft.c @ 7542:a8a8205a9081 libavcodec
split-radix FFT
c is 1.9x faster than previous c (on various x86 cpus), sse is 1.6x faster than previous sse.
author | lorenm |
---|---|
date | Tue, 12 Aug 2008 00:26:58 +0000 |
parents | fc843d00867c |
children | f04ff5a6fb55 |
comparison
equal
deleted
inserted
replaced
7541:570c0c027998 | 7542:a8a8205a9081 |
---|---|
1 /* | 1 /* |
2 * FFT/IFFT transforms | 2 * FFT/IFFT transforms |
3 * Copyright (c) 2008 Loren Merritt | |
3 * Copyright (c) 2002 Fabrice Bellard. | 4 * Copyright (c) 2002 Fabrice Bellard. |
5 * Partly based on libdjbfft by D. J. Bernstein | |
4 * | 6 * |
5 * This file is part of FFmpeg. | 7 * This file is part of FFmpeg. |
6 * | 8 * |
7 * FFmpeg is free software; you can redistribute it and/or | 9 * FFmpeg is free software; you can redistribute it and/or |
8 * modify it under the terms of the GNU Lesser General Public | 10 * modify it under the terms of the GNU Lesser General Public |
24 * FFT/IFFT transforms. | 26 * FFT/IFFT transforms. |
25 */ | 27 */ |
26 | 28 |
27 #include "dsputil.h" | 29 #include "dsputil.h" |
28 | 30 |
31 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */ | |
32 DECLARE_ALIGNED_16(FFTSample, ff_cos_16[8]); | |
33 DECLARE_ALIGNED_16(FFTSample, ff_cos_32[16]); | |
34 DECLARE_ALIGNED_16(FFTSample, ff_cos_64[32]); | |
35 DECLARE_ALIGNED_16(FFTSample, ff_cos_128[64]); | |
36 DECLARE_ALIGNED_16(FFTSample, ff_cos_256[128]); | |
37 DECLARE_ALIGNED_16(FFTSample, ff_cos_512[256]); | |
38 DECLARE_ALIGNED_16(FFTSample, ff_cos_1024[512]); | |
39 DECLARE_ALIGNED_16(FFTSample, ff_cos_2048[1024]); | |
40 DECLARE_ALIGNED_16(FFTSample, ff_cos_4096[2048]); | |
41 DECLARE_ALIGNED_16(FFTSample, ff_cos_8192[4096]); | |
42 DECLARE_ALIGNED_16(FFTSample, ff_cos_16384[8192]); | |
43 DECLARE_ALIGNED_16(FFTSample, ff_cos_32768[16384]); | |
44 DECLARE_ALIGNED_16(FFTSample, ff_cos_65536[32768]); | |
45 static FFTSample *ff_cos_tabs[] = { | |
46 ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256, ff_cos_512, ff_cos_1024, | |
47 ff_cos_2048, ff_cos_4096, ff_cos_8192, ff_cos_16384, ff_cos_32768, ff_cos_65536, | |
48 }; | |
49 | |
50 static int split_radix_permutation(int i, int n, int inverse) | |
51 { | |
52 int m; | |
53 if(n <= 2) return i&1; | |
54 m = n >> 1; | |
55 if(!(i&m)) return split_radix_permutation(i, m, inverse)*2; | |
56 m >>= 1; | |
57 if(inverse == !(i&m)) return split_radix_permutation(i, m, inverse)*4 + 1; | |
58 else return split_radix_permutation(i, m, inverse)*4 - 1; | |
59 } | |
60 | |
29 /** | 61 /** |
30 * The size of the FFT is 2^nbits. If inverse is TRUE, inverse FFT is | 62 * The size of the FFT is 2^nbits. If inverse is TRUE, inverse FFT is |
31 * done | 63 * done |
32 */ | 64 */ |
33 int ff_fft_init(FFTContext *s, int nbits, int inverse) | 65 int ff_fft_init(FFTContext *s, int nbits, int inverse) |
34 { | 66 { |
35 int i, j, m, n; | 67 int i, j, m, n; |
36 float alpha, c1, s1, s2; | 68 float alpha, c1, s1, s2; |
37 int shuffle = 0; | 69 int split_radix = 1; |
38 int av_unused has_vectors; | 70 int av_unused has_vectors; |
39 | 71 |
72 if (nbits < 2 || nbits > 16) | |
73 goto fail; | |
40 s->nbits = nbits; | 74 s->nbits = nbits; |
41 n = 1 << nbits; | 75 n = 1 << nbits; |
42 | 76 |
77 s->tmp_buf = NULL; | |
43 s->exptab = av_malloc((n / 2) * sizeof(FFTComplex)); | 78 s->exptab = av_malloc((n / 2) * sizeof(FFTComplex)); |
44 if (!s->exptab) | 79 if (!s->exptab) |
45 goto fail; | 80 goto fail; |
46 s->revtab = av_malloc(n * sizeof(uint16_t)); | 81 s->revtab = av_malloc(n * sizeof(uint16_t)); |
47 if (!s->revtab) | 82 if (!s->revtab) |
48 goto fail; | 83 goto fail; |
49 s->inverse = inverse; | 84 s->inverse = inverse; |
50 | 85 |
51 s2 = inverse ? 1.0 : -1.0; | 86 s2 = inverse ? 1.0 : -1.0; |
52 | 87 |
53 for(i=0;i<(n/2);i++) { | 88 s->fft_permute = ff_fft_permute_c; |
54 alpha = 2 * M_PI * (float)i / (float)n; | |
55 c1 = cos(alpha); | |
56 s1 = sin(alpha) * s2; | |
57 s->exptab[i].re = c1; | |
58 s->exptab[i].im = s1; | |
59 } | |
60 s->fft_calc = ff_fft_calc_c; | 89 s->fft_calc = ff_fft_calc_c; |
61 s->imdct_calc = ff_imdct_calc; | 90 s->imdct_calc = ff_imdct_calc; |
62 s->imdct_half = ff_imdct_half; | 91 s->imdct_half = ff_imdct_half; |
63 s->exptab1 = NULL; | 92 s->exptab1 = NULL; |
64 | 93 |
65 #ifdef HAVE_MMX | 94 #if defined HAVE_MMX && defined HAVE_YASM |
66 has_vectors = mm_support(); | 95 has_vectors = mm_support(); |
67 shuffle = 1; | 96 if (has_vectors & MM_SSE) { |
68 if (has_vectors & MM_3DNOWEXT) { | 97 /* SSE for P3/P4/K8 */ |
69 /* 3DNowEx for K7/K8 */ | 98 s->imdct_calc = ff_imdct_calc_sse; |
99 s->imdct_half = ff_imdct_half_sse; | |
100 s->fft_permute = ff_fft_permute_sse; | |
101 s->fft_calc = ff_fft_calc_sse; | |
102 } else if (has_vectors & MM_3DNOWEXT) { | |
103 /* 3DNowEx for K7 */ | |
70 s->imdct_calc = ff_imdct_calc_3dn2; | 104 s->imdct_calc = ff_imdct_calc_3dn2; |
71 s->imdct_half = ff_imdct_half_3dn2; | 105 s->imdct_half = ff_imdct_half_3dn2; |
72 s->fft_calc = ff_fft_calc_3dn2; | 106 s->fft_calc = ff_fft_calc_3dn2; |
73 } else if (has_vectors & MM_3DNOW) { | 107 } else if (has_vectors & MM_3DNOW) { |
74 /* 3DNow! for K6-2/3 */ | 108 /* 3DNow! for K6-2/3 */ |
75 s->fft_calc = ff_fft_calc_3dn; | 109 s->fft_calc = ff_fft_calc_3dn; |
76 } else if (has_vectors & MM_SSE) { | |
77 /* SSE for P3/P4 */ | |
78 s->imdct_calc = ff_imdct_calc_sse; | |
79 s->imdct_half = ff_imdct_half_sse; | |
80 s->fft_calc = ff_fft_calc_sse; | |
81 } else { | |
82 shuffle = 0; | |
83 } | 110 } |
84 #elif defined HAVE_ALTIVEC && !defined ALTIVEC_USE_REFERENCE_C_CODE | 111 #elif defined HAVE_ALTIVEC && !defined ALTIVEC_USE_REFERENCE_C_CODE |
85 has_vectors = mm_support(); | 112 has_vectors = mm_support(); |
86 if (has_vectors & MM_ALTIVEC) { | 113 if (has_vectors & MM_ALTIVEC) { |
87 s->fft_calc = ff_fft_calc_altivec; | 114 s->fft_calc = ff_fft_calc_altivec; |
88 shuffle = 1; | 115 split_radix = 0; |
89 } | 116 } |
90 #endif | 117 #endif |
91 | 118 |
92 /* compute constant table for HAVE_SSE version */ | 119 if (split_radix) { |
93 if (shuffle) { | 120 for(j=4; j<=nbits; j++) { |
121 int m = 1<<j; | |
122 double freq = 2*M_PI/m; | |
123 FFTSample *tab = ff_cos_tabs[j-4]; | |
124 for(i=0; i<=m/4; i++) | |
125 tab[i] = cos(i*freq); | |
126 for(i=1; i<m/4; i++) | |
127 tab[m/2-i] = tab[i]; | |
128 } | |
129 for(i=0; i<n; i++) | |
130 s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = i; | |
131 s->tmp_buf = av_malloc(n * sizeof(FFTComplex)); | |
132 } else { | |
94 int np, nblocks, np2, l; | 133 int np, nblocks, np2, l; |
95 FFTComplex *q; | 134 FFTComplex *q; |
135 | |
136 for(i=0; i<(n/2); i++) { | |
137 alpha = 2 * M_PI * (float)i / (float)n; | |
138 c1 = cos(alpha); | |
139 s1 = sin(alpha) * s2; | |
140 s->exptab[i].re = c1; | |
141 s->exptab[i].im = s1; | |
142 } | |
96 | 143 |
97 np = 1 << nbits; | 144 np = 1 << nbits; |
98 nblocks = np >> 3; | 145 nblocks = np >> 3; |
99 np2 = np >> 1; | 146 np2 = np >> 1; |
100 s->exptab1 = av_malloc(np * 2 * sizeof(FFTComplex)); | 147 s->exptab1 = av_malloc(np * 2 * sizeof(FFTComplex)); |
114 q++; | 161 q++; |
115 } | 162 } |
116 nblocks = nblocks >> 1; | 163 nblocks = nblocks >> 1; |
117 } while (nblocks != 0); | 164 } while (nblocks != 0); |
118 av_freep(&s->exptab); | 165 av_freep(&s->exptab); |
119 } | |
120 | 166 |
121 /* compute bit reverse table */ | 167 /* compute bit reverse table */ |
122 | 168 |
123 for(i=0;i<n;i++) { | 169 for(i=0;i<n;i++) { |
124 m=0; | 170 m=0; |
125 for(j=0;j<nbits;j++) { | 171 for(j=0;j<nbits;j++) { |
126 m |= ((i >> j) & 1) << (nbits-j-1); | 172 m |= ((i >> j) & 1) << (nbits-j-1); |
127 } | 173 } |
128 s->revtab[i]=m; | 174 s->revtab[i]=m; |
129 } | 175 } |
176 } | |
177 | |
130 return 0; | 178 return 0; |
131 fail: | 179 fail: |
132 av_freep(&s->revtab); | 180 av_freep(&s->revtab); |
133 av_freep(&s->exptab); | 181 av_freep(&s->exptab); |
134 av_freep(&s->exptab1); | 182 av_freep(&s->exptab1); |
183 av_freep(&s->tmp_buf); | |
135 return -1; | 184 return -1; |
136 } | |
137 | |
138 /* butter fly op */ | |
139 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \ | |
140 {\ | |
141 FFTSample ax, ay, bx, by;\ | |
142 bx=pre1;\ | |
143 by=pim1;\ | |
144 ax=qre1;\ | |
145 ay=qim1;\ | |
146 pre = (bx + ax);\ | |
147 pim = (by + ay);\ | |
148 qre = (bx - ax);\ | |
149 qim = (by - ay);\ | |
150 } | |
151 | |
152 #define MUL16(a,b) ((a) * (b)) | |
153 | |
154 #define CMUL(pre, pim, are, aim, bre, bim) \ | |
155 {\ | |
156 pre = (MUL16(are, bre) - MUL16(aim, bim));\ | |
157 pim = (MUL16(are, bim) + MUL16(bre, aim));\ | |
158 } | |
159 | |
160 /** | |
161 * Do a complex FFT with the parameters defined in ff_fft_init(). The | |
162 * input data must be permuted before with s->revtab table. No | |
163 * 1.0/sqrt(n) normalization is done. | |
164 */ | |
165 void ff_fft_calc_c(FFTContext *s, FFTComplex *z) | |
166 { | |
167 int ln = s->nbits; | |
168 int j, np, np2; | |
169 int nblocks, nloops; | |
170 register FFTComplex *p, *q; | |
171 FFTComplex *exptab = s->exptab; | |
172 int l; | |
173 FFTSample tmp_re, tmp_im; | |
174 | |
175 np = 1 << ln; | |
176 | |
177 /* pass 0 */ | |
178 | |
179 p=&z[0]; | |
180 j=(np >> 1); | |
181 do { | |
182 BF(p[0].re, p[0].im, p[1].re, p[1].im, | |
183 p[0].re, p[0].im, p[1].re, p[1].im); | |
184 p+=2; | |
185 } while (--j != 0); | |
186 | |
187 /* pass 1 */ | |
188 | |
189 | |
190 p=&z[0]; | |
191 j=np >> 2; | |
192 if (s->inverse) { | |
193 do { | |
194 BF(p[0].re, p[0].im, p[2].re, p[2].im, | |
195 p[0].re, p[0].im, p[2].re, p[2].im); | |
196 BF(p[1].re, p[1].im, p[3].re, p[3].im, | |
197 p[1].re, p[1].im, -p[3].im, p[3].re); | |
198 p+=4; | |
199 } while (--j != 0); | |
200 } else { | |
201 do { | |
202 BF(p[0].re, p[0].im, p[2].re, p[2].im, | |
203 p[0].re, p[0].im, p[2].re, p[2].im); | |
204 BF(p[1].re, p[1].im, p[3].re, p[3].im, | |
205 p[1].re, p[1].im, p[3].im, -p[3].re); | |
206 p+=4; | |
207 } while (--j != 0); | |
208 } | |
209 /* pass 2 .. ln-1 */ | |
210 | |
211 nblocks = np >> 3; | |
212 nloops = 1 << 2; | |
213 np2 = np >> 1; | |
214 do { | |
215 p = z; | |
216 q = z + nloops; | |
217 for (j = 0; j < nblocks; ++j) { | |
218 BF(p->re, p->im, q->re, q->im, | |
219 p->re, p->im, q->re, q->im); | |
220 | |
221 p++; | |
222 q++; | |
223 for(l = nblocks; l < np2; l += nblocks) { | |
224 CMUL(tmp_re, tmp_im, exptab[l].re, exptab[l].im, q->re, q->im); | |
225 BF(p->re, p->im, q->re, q->im, | |
226 p->re, p->im, tmp_re, tmp_im); | |
227 p++; | |
228 q++; | |
229 } | |
230 | |
231 p += nloops; | |
232 q += nloops; | |
233 } | |
234 nblocks = nblocks >> 1; | |
235 nloops = nloops << 1; | |
236 } while (nblocks != 0); | |
237 } | 185 } |
238 | 186 |
239 /** | 187 /** |
240 * Do the permutation needed BEFORE calling ff_fft_calc() | 188 * Do the permutation needed BEFORE calling ff_fft_calc() |
241 */ | 189 */ |
242 void ff_fft_permute(FFTContext *s, FFTComplex *z) | 190 void ff_fft_permute_c(FFTContext *s, FFTComplex *z) |
243 { | 191 { |
244 int j, k, np; | 192 int j, k, np; |
245 FFTComplex tmp; | 193 FFTComplex tmp; |
246 const uint16_t *revtab = s->revtab; | 194 const uint16_t *revtab = s->revtab; |
195 np = 1 << s->nbits; | |
196 | |
197 if (s->tmp_buf) { | |
198 /* TODO: handle split-radix permute in a more optimal way, probably in-place */ | |
199 for(j=0;j<np;j++) s->tmp_buf[revtab[j]] = z[j]; | |
200 memcpy(z, s->tmp_buf, np * sizeof(FFTComplex)); | |
201 return; | |
202 } | |
247 | 203 |
248 /* reverse */ | 204 /* reverse */ |
249 np = 1 << s->nbits; | |
250 for(j=0;j<np;j++) { | 205 for(j=0;j<np;j++) { |
251 k = revtab[j]; | 206 k = revtab[j]; |
252 if (k < j) { | 207 if (k < j) { |
253 tmp = z[k]; | 208 tmp = z[k]; |
254 z[k] = z[j]; | 209 z[k] = z[j]; |
260 void ff_fft_end(FFTContext *s) | 215 void ff_fft_end(FFTContext *s) |
261 { | 216 { |
262 av_freep(&s->revtab); | 217 av_freep(&s->revtab); |
263 av_freep(&s->exptab); | 218 av_freep(&s->exptab); |
264 av_freep(&s->exptab1); | 219 av_freep(&s->exptab1); |
265 } | 220 av_freep(&s->tmp_buf); |
266 | 221 } |
222 | |
223 #define sqrthalf (float)M_SQRT1_2 | |
224 | |
225 #define BF(x,y,a,b) {\ | |
226 x = a - b;\ | |
227 y = a + b;\ | |
228 } | |
229 | |
230 #define BUTTERFLIES(a0,a1,a2,a3) {\ | |
231 BF(t3, t5, t5, t1);\ | |
232 BF(a2.re, a0.re, a0.re, t5);\ | |
233 BF(a3.im, a1.im, a1.im, t3);\ | |
234 BF(t4, t6, t2, t6);\ | |
235 BF(a3.re, a1.re, a1.re, t4);\ | |
236 BF(a2.im, a0.im, a0.im, t6);\ | |
237 } | |
238 | |
239 // force loading all the inputs before storing any. | |
240 // this is slightly slower for small data, but avoids store->load aliasing | |
241 // for addresses separated by large powers of 2. | |
242 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\ | |
243 FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\ | |
244 BF(t3, t5, t5, t1);\ | |
245 BF(a2.re, a0.re, r0, t5);\ | |
246 BF(a3.im, a1.im, i1, t3);\ | |
247 BF(t4, t6, t2, t6);\ | |
248 BF(a3.re, a1.re, r1, t4);\ | |
249 BF(a2.im, a0.im, i0, t6);\ | |
250 } | |
251 | |
252 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\ | |
253 t1 = a2.re * wre + a2.im * wim;\ | |
254 t2 = a2.im * wre - a2.re * wim;\ | |
255 t5 = a3.re * wre - a3.im * wim;\ | |
256 t6 = a3.im * wre + a3.re * wim;\ | |
257 BUTTERFLIES(a0,a1,a2,a3)\ | |
258 } | |
259 | |
260 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\ | |
261 t1 = a2.re;\ | |
262 t2 = a2.im;\ | |
263 t5 = a3.re;\ | |
264 t6 = a3.im;\ | |
265 BUTTERFLIES(a0,a1,a2,a3)\ | |
266 } | |
267 | |
268 /* z[0...8n-1], w[1...2n-1] */ | |
269 #define PASS(name)\ | |
270 static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\ | |
271 {\ | |
272 FFTSample t1, t2, t3, t4, t5, t6;\ | |
273 int o1 = 2*n;\ | |
274 int o2 = 4*n;\ | |
275 int o3 = 6*n;\ | |
276 const FFTSample *wim = wre+o1;\ | |
277 n--;\ | |
278 \ | |
279 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\ | |
280 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\ | |
281 do {\ | |
282 z += 2;\ | |
283 wre += 2;\ | |
284 wim -= 2;\ | |
285 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\ | |
286 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\ | |
287 } while(--n);\ | |
288 } | |
289 | |
290 PASS(pass) | |
291 #undef BUTTERFLIES | |
292 #define BUTTERFLIES BUTTERFLIES_BIG | |
293 PASS(pass_big) | |
294 | |
295 #define DECL_FFT(n,n2,n4)\ | |
296 static void fft##n(FFTComplex *z)\ | |
297 {\ | |
298 fft##n2(z);\ | |
299 fft##n4(z+n4*2);\ | |
300 fft##n4(z+n4*3);\ | |
301 pass(z,ff_cos_##n,n4/2);\ | |
302 } | |
303 | |
304 static void fft4(FFTComplex *z) | |
305 { | |
306 FFTSample t1, t2, t3, t4, t5, t6, t7, t8; | |
307 | |
308 BF(t3, t1, z[0].re, z[1].re); | |
309 BF(t8, t6, z[3].re, z[2].re); | |
310 BF(z[2].re, z[0].re, t1, t6); | |
311 BF(t4, t2, z[0].im, z[1].im); | |
312 BF(t7, t5, z[2].im, z[3].im); | |
313 BF(z[3].im, z[1].im, t4, t8); | |
314 BF(z[3].re, z[1].re, t3, t7); | |
315 BF(z[2].im, z[0].im, t2, t5); | |
316 } | |
317 | |
318 static void fft8(FFTComplex *z) | |
319 { | |
320 FFTSample t1, t2, t3, t4, t5, t6, t7, t8; | |
321 | |
322 fft4(z); | |
323 | |
324 BF(t1, z[5].re, z[4].re, -z[5].re); | |
325 BF(t2, z[5].im, z[4].im, -z[5].im); | |
326 BF(t3, z[7].re, z[6].re, -z[7].re); | |
327 BF(t4, z[7].im, z[6].im, -z[7].im); | |
328 BF(t8, t1, t3, t1); | |
329 BF(t7, t2, t2, t4); | |
330 BF(z[4].re, z[0].re, z[0].re, t1); | |
331 BF(z[4].im, z[0].im, z[0].im, t2); | |
332 BF(z[6].re, z[2].re, z[2].re, t7); | |
333 BF(z[6].im, z[2].im, z[2].im, t8); | |
334 | |
335 TRANSFORM(z[1],z[3],z[5],z[7],sqrthalf,sqrthalf); | |
336 } | |
337 | |
338 #ifndef CONFIG_SMALL | |
339 static void fft16(FFTComplex *z) | |
340 { | |
341 FFTSample t1, t2, t3, t4, t5, t6; | |
342 | |
343 fft8(z); | |
344 fft4(z+8); | |
345 fft4(z+12); | |
346 | |
347 TRANSFORM_ZERO(z[0],z[4],z[8],z[12]); | |
348 TRANSFORM(z[2],z[6],z[10],z[14],sqrthalf,sqrthalf); | |
349 TRANSFORM(z[1],z[5],z[9],z[13],ff_cos_16[1],ff_cos_16[3]); | |
350 TRANSFORM(z[3],z[7],z[11],z[15],ff_cos_16[3],ff_cos_16[1]); | |
351 } | |
352 #else | |
353 DECL_FFT(16,8,4) | |
354 #endif | |
355 DECL_FFT(32,16,8) | |
356 DECL_FFT(64,32,16) | |
357 DECL_FFT(128,64,32) | |
358 DECL_FFT(256,128,64) | |
359 DECL_FFT(512,256,128) | |
360 #ifndef CONFIG_SMALL | |
361 #define pass pass_big | |
362 #endif | |
363 DECL_FFT(1024,512,256) | |
364 DECL_FFT(2048,1024,512) | |
365 DECL_FFT(4096,2048,1024) | |
366 DECL_FFT(8192,4096,2048) | |
367 DECL_FFT(16384,8192,4096) | |
368 DECL_FFT(32768,16384,8192) | |
369 DECL_FFT(65536,32768,16384) | |
370 | |
371 static void (*fft_dispatch[])(FFTComplex*) = { | |
372 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024, | |
373 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536, | |
374 }; | |
375 | |
376 /** | |
377 * Do a complex FFT with the parameters defined in ff_fft_init(). The | |
378 * input data must be permuted before with s->revtab table. No | |
379 * 1.0/sqrt(n) normalization is done. | |
380 */ | |
381 void ff_fft_calc_c(FFTContext *s, FFTComplex *z) | |
382 { | |
383 fft_dispatch[s->nbits-2](z); | |
384 } | |
385 |