Fixed-point FFT and MDCT
[libav.git] / libavcodec / fft.c
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
24 /**
25 * @file
26 * FFT/IFFT transforms.
27 */
28
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
82 av_cold void ff_init_ff_cos_tabs(int index)
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 av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
97 {
98 int i, j, n;
99
100 if (nbits < 2 || nbits > 16)
101 goto fail;
102 s->nbits = nbits;
103 n = 1 << nbits;
104
105 s->revtab = av_malloc(n * sizeof(uint16_t));
106 if (!s->revtab)
107 goto fail;
108 s->tmp_buf = av_malloc(n * sizeof(FFTComplex));
109 if (!s->tmp_buf)
110 goto fail;
111 s->inverse = inverse;
112 s->fft_permutation = FF_FFT_PERM_DEFAULT;
113
114 s->fft_permute = ff_fft_permute_c;
115 s->fft_calc = ff_fft_calc_c;
116 #if CONFIG_MDCT
117 s->imdct_calc = ff_imdct_calc_c;
118 s->imdct_half = ff_imdct_half_c;
119 s->mdct_calc = ff_mdct_calc_c;
120 #endif
121
122 #if CONFIG_FFT_FLOAT
123 if (ARCH_ARM) ff_fft_init_arm(s);
124 if (HAVE_ALTIVEC) ff_fft_init_altivec(s);
125 if (HAVE_MMX) ff_fft_init_mmx(s);
126 #endif
127
128 for(j=4; j<=nbits; j++) {
129 ff_init_ff_cos_tabs(j);
130 }
131 for(i=0; i<n; i++) {
132 int j = i;
133 if (s->fft_permutation == FF_FFT_PERM_SWAP_LSBS)
134 j = (j&~3) | ((j>>1)&1) | ((j<<1)&2);
135 s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = j;
136 }
137
138 return 0;
139 fail:
140 av_freep(&s->revtab);
141 av_freep(&s->tmp_buf);
142 return -1;
143 }
144
145 static void ff_fft_permute_c(FFTContext *s, FFTComplex *z)
146 {
147 int j, np;
148 const uint16_t *revtab = s->revtab;
149 np = 1 << s->nbits;
150 /* TODO: handle split-radix permute in a more optimal way, probably in-place */
151 for(j=0;j<np;j++) s->tmp_buf[revtab[j]] = z[j];
152 memcpy(z, s->tmp_buf, np * sizeof(FFTComplex));
153 }
154
155 av_cold void ff_fft_end(FFTContext *s)
156 {
157 av_freep(&s->revtab);
158 av_freep(&s->tmp_buf);
159 }
160
161 #define BUTTERFLIES(a0,a1,a2,a3) {\
162 BF(t3, t5, t5, t1);\
163 BF(a2.re, a0.re, a0.re, t5);\
164 BF(a3.im, a1.im, a1.im, t3);\
165 BF(t4, t6, t2, t6);\
166 BF(a3.re, a1.re, a1.re, t4);\
167 BF(a2.im, a0.im, a0.im, t6);\
168 }
169
170 // force loading all the inputs before storing any.
171 // this is slightly slower for small data, but avoids store->load aliasing
172 // for addresses separated by large powers of 2.
173 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
174 FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
175 BF(t3, t5, t5, t1);\
176 BF(a2.re, a0.re, r0, t5);\
177 BF(a3.im, a1.im, i1, t3);\
178 BF(t4, t6, t2, t6);\
179 BF(a3.re, a1.re, r1, t4);\
180 BF(a2.im, a0.im, i0, t6);\
181 }
182
183 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
184 CMUL(t1, t2, a2.re, a2.im, wre, -wim);\
185 CMUL(t5, t6, a3.re, a3.im, wre, wim);\
186 BUTTERFLIES(a0,a1,a2,a3)\
187 }
188
189 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
190 t1 = a2.re;\
191 t2 = a2.im;\
192 t5 = a3.re;\
193 t6 = a3.im;\
194 BUTTERFLIES(a0,a1,a2,a3)\
195 }
196
197 /* z[0...8n-1], w[1...2n-1] */
198 #define PASS(name)\
199 static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\
200 {\
201 FFTDouble t1, t2, t3, t4, t5, t6;\
202 int o1 = 2*n;\
203 int o2 = 4*n;\
204 int o3 = 6*n;\
205 const FFTSample *wim = wre+o1;\
206 n--;\
207 \
208 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
209 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
210 do {\
211 z += 2;\
212 wre += 2;\
213 wim -= 2;\
214 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
215 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
216 } while(--n);\
217 }
218
219 PASS(pass)
220 #undef BUTTERFLIES
221 #define BUTTERFLIES BUTTERFLIES_BIG
222 PASS(pass_big)
223
224 #define DECL_FFT(n,n2,n4)\
225 static void fft##n(FFTComplex *z)\
226 {\
227 fft##n2(z);\
228 fft##n4(z+n4*2);\
229 fft##n4(z+n4*3);\
230 pass(z,FFT_NAME(ff_cos_##n),n4/2);\
231 }
232
233 static void fft4(FFTComplex *z)
234 {
235 FFTDouble t1, t2, t3, t4, t5, t6, t7, t8;
236
237 BF(t3, t1, z[0].re, z[1].re);
238 BF(t8, t6, z[3].re, z[2].re);
239 BF(z[2].re, z[0].re, t1, t6);
240 BF(t4, t2, z[0].im, z[1].im);
241 BF(t7, t5, z[2].im, z[3].im);
242 BF(z[3].im, z[1].im, t4, t8);
243 BF(z[3].re, z[1].re, t3, t7);
244 BF(z[2].im, z[0].im, t2, t5);
245 }
246
247 static void fft8(FFTComplex *z)
248 {
249 FFTDouble t1, t2, t3, t4, t5, t6, t7, t8;
250
251 fft4(z);
252
253 BF(t1, z[5].re, z[4].re, -z[5].re);
254 BF(t2, z[5].im, z[4].im, -z[5].im);
255 BF(t3, z[7].re, z[6].re, -z[7].re);
256 BF(t4, z[7].im, z[6].im, -z[7].im);
257 BF(t8, t1, t3, t1);
258 BF(t7, t2, t2, t4);
259 BF(z[4].re, z[0].re, z[0].re, t1);
260 BF(z[4].im, z[0].im, z[0].im, t2);
261 BF(z[6].re, z[2].re, z[2].re, t7);
262 BF(z[6].im, z[2].im, z[2].im, t8);
263
264 TRANSFORM(z[1],z[3],z[5],z[7],sqrthalf,sqrthalf);
265 }
266
267 #if !CONFIG_SMALL
268 static void fft16(FFTComplex *z)
269 {
270 FFTDouble t1, t2, t3, t4, t5, t6;
271 FFTSample cos_16_1 = FFT_NAME(ff_cos_16)[1];
272 FFTSample cos_16_3 = FFT_NAME(ff_cos_16)[3];
273
274 fft8(z);
275 fft4(z+8);
276 fft4(z+12);
277
278 TRANSFORM_ZERO(z[0],z[4],z[8],z[12]);
279 TRANSFORM(z[2],z[6],z[10],z[14],sqrthalf,sqrthalf);
280 TRANSFORM(z[1],z[5],z[9],z[13],cos_16_1,cos_16_3);
281 TRANSFORM(z[3],z[7],z[11],z[15],cos_16_3,cos_16_1);
282 }
283 #else
284 DECL_FFT(16,8,4)
285 #endif
286 DECL_FFT(32,16,8)
287 DECL_FFT(64,32,16)
288 DECL_FFT(128,64,32)
289 DECL_FFT(256,128,64)
290 DECL_FFT(512,256,128)
291 #if !CONFIG_SMALL
292 #define pass pass_big
293 #endif
294 DECL_FFT(1024,512,256)
295 DECL_FFT(2048,1024,512)
296 DECL_FFT(4096,2048,1024)
297 DECL_FFT(8192,4096,2048)
298 DECL_FFT(16384,8192,4096)
299 DECL_FFT(32768,16384,8192)
300 DECL_FFT(65536,32768,16384)
301
302 static void (* const fft_dispatch[])(FFTComplex*) = {
303 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
304 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536,
305 };
306
307 static void ff_fft_calc_c(FFTContext *s, FFTComplex *z)
308 {
309 fft_dispatch[s->nbits-2](z);
310 }
311