Mercurial > libavcodec.hg
annotate fft.c @ 10492:63910f7ba293 libavcodec
Pad ff_cos_tabs and ff_sin_tabs so that index n points to the table for n bits.
While this "wastes" up to 2x32 bytes it makes the code slightly simpler and
less confusing.
author | reimar |
---|---|
date | Wed, 04 Nov 2009 17:30:23 +0000 |
parents | 57acce8b1380 |
children | 74b0c1a0851e |
rev | line source |
---|---|
781 | 1 /* |
2 * FFT/IFFT transforms | |
7542 | 3 * Copyright (c) 2008 Loren Merritt |
8629
04423b2f6e0b
cosmetics: Remove pointless period after copyright statement non-sentences.
diego
parents:
8590
diff
changeset
|
4 * Copyright (c) 2002 Fabrice Bellard |
7542 | 5 * Partly based on libdjbfft by D. J. Bernstein |
781 | 6 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
7 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
8 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
9 * FFmpeg is free software; you can redistribute it and/or |
781 | 10 * modify it under the terms of the GNU Lesser General Public |
11 * License as published by the Free Software Foundation; either | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
12 * version 2.1 of the License, or (at your option) any later version. |
781 | 13 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
14 * FFmpeg is distributed in the hope that it will be useful, |
781 | 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 | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
20 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
781 | 22 */ |
1106 | 23 |
24 /** | |
8718
e9d9d946f213
Use full internal pathname in doxygen @file directives.
diego
parents:
8694
diff
changeset
|
25 * @file libavcodec/fft.c |
1106 | 26 * FFT/IFFT transforms. |
27 */ | |
28 | |
781 | 29 #include "dsputil.h" |
30 | |
10407
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
31 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */ |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
32 #if !CONFIG_HARDCODED_TABLES |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
33 COSTABLE(16); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
34 COSTABLE(32); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
35 COSTABLE(64); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
36 COSTABLE(128); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
37 COSTABLE(256); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
38 COSTABLE(512); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
39 COSTABLE(1024); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
40 COSTABLE(2048); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
41 COSTABLE(4096); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
42 COSTABLE(8192); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
43 COSTABLE(16384); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
44 COSTABLE(32768); |
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
45 COSTABLE(65536); |
10400
866dffa620d1
Use hardcoded instead of runtime-calculated ff_cos_* tables if
reimar
parents:
10391
diff
changeset
|
46 #endif |
10407
57acce8b1380
Move/add COSTABLE/SINTABLE macros to dsputil to add extern definitions
reimar
parents:
10400
diff
changeset
|
47 COSTABLE_CONST FFTSample * const ff_cos_tabs[] = { |
10492
63910f7ba293
Pad ff_cos_tabs and ff_sin_tabs so that index n points to the table for n bits.
reimar
parents:
10407
diff
changeset
|
48 NULL, NULL, NULL, NULL, |
7542 | 49 ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256, ff_cos_512, ff_cos_1024, |
50 ff_cos_2048, ff_cos_4096, ff_cos_8192, ff_cos_16384, ff_cos_32768, ff_cos_65536, | |
51 }; | |
52 | |
53 static int split_radix_permutation(int i, int n, int inverse) | |
54 { | |
55 int m; | |
56 if(n <= 2) return i&1; | |
57 m = n >> 1; | |
58 if(!(i&m)) return split_radix_permutation(i, m, inverse)*2; | |
59 m >>= 1; | |
60 if(inverse == !(i&m)) return split_radix_permutation(i, m, inverse)*4 + 1; | |
61 else return split_radix_permutation(i, m, inverse)*4 - 1; | |
62 } | |
63 | |
8637 | 64 av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse) |
781 | 65 { |
66 int i, j, m, n; | |
67 float alpha, c1, s1, s2; | |
6503 | 68 int av_unused has_vectors; |
2967 | 69 |
7542 | 70 if (nbits < 2 || nbits > 16) |
71 goto fail; | |
781 | 72 s->nbits = nbits; |
73 n = 1 << nbits; | |
74 | |
7542 | 75 s->tmp_buf = NULL; |
8974 | 76 s->exptab = av_malloc((n / 2) * sizeof(FFTComplex)); |
781 | 77 if (!s->exptab) |
78 goto fail; | |
79 s->revtab = av_malloc(n * sizeof(uint16_t)); | |
80 if (!s->revtab) | |
81 goto fail; | |
82 s->inverse = inverse; | |
83 | |
84 s2 = inverse ? 1.0 : -1.0; | |
2967 | 85 |
7542 | 86 s->fft_permute = ff_fft_permute_c; |
8974 | 87 s->fft_calc = ff_fft_calc_c; |
88 s->imdct_calc = ff_imdct_calc_c; | |
89 s->imdct_half = ff_imdct_half_c; | |
10161 | 90 s->mdct_calc = ff_mdct_calc_c; |
8974 | 91 s->exptab1 = NULL; |
10175
5cf49858179a
Move per-arch fft init bits into the corresponding subdirs
mru
parents:
10172
diff
changeset
|
92 s->split_radix = 1; |
781 | 93 |
10175
5cf49858179a
Move per-arch fft init bits into the corresponding subdirs
mru
parents:
10172
diff
changeset
|
94 if (ARCH_ARM) ff_fft_init_arm(s); |
5cf49858179a
Move per-arch fft init bits into the corresponding subdirs
mru
parents:
10172
diff
changeset
|
95 if (HAVE_ALTIVEC) ff_fft_init_altivec(s); |
5cf49858179a
Move per-arch fft init bits into the corresponding subdirs
mru
parents:
10172
diff
changeset
|
96 if (HAVE_MMX) ff_fft_init_mmx(s); |
781 | 97 |
10175
5cf49858179a
Move per-arch fft init bits into the corresponding subdirs
mru
parents:
10172
diff
changeset
|
98 if (s->split_radix) { |
10400
866dffa620d1
Use hardcoded instead of runtime-calculated ff_cos_* tables if
reimar
parents:
10391
diff
changeset
|
99 #if !CONFIG_HARDCODED_TABLES |
7542 | 100 for(j=4; j<=nbits; j++) { |
101 int m = 1<<j; | |
102 double freq = 2*M_PI/m; | |
10492
63910f7ba293
Pad ff_cos_tabs and ff_sin_tabs so that index n points to the table for n bits.
reimar
parents:
10407
diff
changeset
|
103 FFTSample *tab = ff_cos_tabs[j]; |
7542 | 104 for(i=0; i<=m/4; i++) |
105 tab[i] = cos(i*freq); | |
106 for(i=1; i<m/4; i++) | |
107 tab[m/2-i] = tab[i]; | |
108 } | |
10400
866dffa620d1
Use hardcoded instead of runtime-calculated ff_cos_* tables if
reimar
parents:
10391
diff
changeset
|
109 #endif |
7542 | 110 for(i=0; i<n; i++) |
10172 | 111 s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = i; |
7542 | 112 s->tmp_buf = av_malloc(n * sizeof(FFTComplex)); |
113 } else { | |
6504 | 114 int np, nblocks, np2, l; |
115 FFTComplex *q; | |
2967 | 116 |
7542 | 117 for(i=0; i<(n/2); i++) { |
118 alpha = 2 * M_PI * (float)i / (float)n; | |
119 c1 = cos(alpha); | |
120 s1 = sin(alpha) * s2; | |
121 s->exptab[i].re = c1; | |
122 s->exptab[i].im = s1; | |
123 } | |
124 | |
6504 | 125 np = 1 << nbits; |
126 nblocks = np >> 3; | |
127 np2 = np >> 1; | |
128 s->exptab1 = av_malloc(np * 2 * sizeof(FFTComplex)); | |
129 if (!s->exptab1) | |
130 goto fail; | |
131 q = s->exptab1; | |
132 do { | |
133 for(l = 0; l < np2; l += 2 * nblocks) { | |
134 *q++ = s->exptab[l]; | |
135 *q++ = s->exptab[l + nblocks]; | |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
136 |
6504 | 137 q->re = -s->exptab[l].im; |
138 q->im = s->exptab[l].re; | |
139 q++; | |
140 q->re = -s->exptab[l + nblocks].im; | |
141 q->im = s->exptab[l + nblocks].re; | |
142 q++; | |
143 } | |
144 nblocks = nblocks >> 1; | |
145 } while (nblocks != 0); | |
146 av_freep(&s->exptab); | |
781 | 147 |
7543 | 148 /* compute bit reverse table */ |
149 for(i=0;i<n;i++) { | |
150 m=0; | |
151 for(j=0;j<nbits;j++) { | |
152 m |= ((i >> j) & 1) << (nbits-j-1); | |
153 } | |
154 s->revtab[i]=m; | |
781 | 155 } |
7542 | 156 } |
157 | |
781 | 158 return 0; |
159 fail: | |
160 av_freep(&s->revtab); | |
161 av_freep(&s->exptab); | |
162 av_freep(&s->exptab1); | |
7542 | 163 av_freep(&s->tmp_buf); |
781 | 164 return -1; |
165 } | |
166 | |
7542 | 167 void ff_fft_permute_c(FFTContext *s, FFTComplex *z) |
781 | 168 { |
169 int j, k, np; | |
170 FFTComplex tmp; | |
171 const uint16_t *revtab = s->revtab; | |
7542 | 172 np = 1 << s->nbits; |
173 | |
174 if (s->tmp_buf) { | |
175 /* TODO: handle split-radix permute in a more optimal way, probably in-place */ | |
176 for(j=0;j<np;j++) s->tmp_buf[revtab[j]] = z[j]; | |
177 memcpy(z, s->tmp_buf, np * sizeof(FFTComplex)); | |
178 return; | |
179 } | |
2967 | 180 |
781 | 181 /* reverse */ |
182 for(j=0;j<np;j++) { | |
183 k = revtab[j]; | |
184 if (k < j) { | |
185 tmp = z[k]; | |
186 z[k] = z[j]; | |
187 z[j] = tmp; | |
188 } | |
189 } | |
190 } | |
191 | |
8687 | 192 av_cold void ff_fft_end(FFTContext *s) |
781 | 193 { |
194 av_freep(&s->revtab); | |
195 av_freep(&s->exptab); | |
196 av_freep(&s->exptab1); | |
7542 | 197 av_freep(&s->tmp_buf); |
198 } | |
199 | |
200 #define sqrthalf (float)M_SQRT1_2 | |
201 | |
202 #define BF(x,y,a,b) {\ | |
203 x = a - b;\ | |
204 y = a + b;\ | |
205 } | |
206 | |
207 #define BUTTERFLIES(a0,a1,a2,a3) {\ | |
208 BF(t3, t5, t5, t1);\ | |
209 BF(a2.re, a0.re, a0.re, t5);\ | |
210 BF(a3.im, a1.im, a1.im, t3);\ | |
211 BF(t4, t6, t2, t6);\ | |
212 BF(a3.re, a1.re, a1.re, t4);\ | |
213 BF(a2.im, a0.im, a0.im, t6);\ | |
214 } | |
215 | |
216 // force loading all the inputs before storing any. | |
217 // this is slightly slower for small data, but avoids store->load aliasing | |
218 // for addresses separated by large powers of 2. | |
219 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\ | |
220 FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\ | |
221 BF(t3, t5, t5, t1);\ | |
222 BF(a2.re, a0.re, r0, t5);\ | |
223 BF(a3.im, a1.im, i1, t3);\ | |
224 BF(t4, t6, t2, t6);\ | |
225 BF(a3.re, a1.re, r1, t4);\ | |
226 BF(a2.im, a0.im, i0, t6);\ | |
227 } | |
228 | |
229 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\ | |
230 t1 = a2.re * wre + a2.im * wim;\ | |
231 t2 = a2.im * wre - a2.re * wim;\ | |
232 t5 = a3.re * wre - a3.im * wim;\ | |
233 t6 = a3.im * wre + a3.re * wim;\ | |
234 BUTTERFLIES(a0,a1,a2,a3)\ | |
235 } | |
236 | |
237 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\ | |
238 t1 = a2.re;\ | |
239 t2 = a2.im;\ | |
240 t5 = a3.re;\ | |
241 t6 = a3.im;\ | |
242 BUTTERFLIES(a0,a1,a2,a3)\ | |
243 } | |
244 | |
245 /* z[0...8n-1], w[1...2n-1] */ | |
246 #define PASS(name)\ | |
247 static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\ | |
248 {\ | |
249 FFTSample t1, t2, t3, t4, t5, t6;\ | |
250 int o1 = 2*n;\ | |
251 int o2 = 4*n;\ | |
252 int o3 = 6*n;\ | |
253 const FFTSample *wim = wre+o1;\ | |
254 n--;\ | |
255 \ | |
256 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\ | |
257 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\ | |
258 do {\ | |
259 z += 2;\ | |
260 wre += 2;\ | |
261 wim -= 2;\ | |
262 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\ | |
263 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\ | |
264 } while(--n);\ | |
781 | 265 } |
266 | |
7542 | 267 PASS(pass) |
268 #undef BUTTERFLIES | |
269 #define BUTTERFLIES BUTTERFLIES_BIG | |
270 PASS(pass_big) | |
271 | |
272 #define DECL_FFT(n,n2,n4)\ | |
273 static void fft##n(FFTComplex *z)\ | |
274 {\ | |
275 fft##n2(z);\ | |
276 fft##n4(z+n4*2);\ | |
277 fft##n4(z+n4*3);\ | |
278 pass(z,ff_cos_##n,n4/2);\ | |
279 } | |
280 | |
281 static void fft4(FFTComplex *z) | |
282 { | |
283 FFTSample t1, t2, t3, t4, t5, t6, t7, t8; | |
284 | |
285 BF(t3, t1, z[0].re, z[1].re); | |
286 BF(t8, t6, z[3].re, z[2].re); | |
287 BF(z[2].re, z[0].re, t1, t6); | |
288 BF(t4, t2, z[0].im, z[1].im); | |
289 BF(t7, t5, z[2].im, z[3].im); | |
290 BF(z[3].im, z[1].im, t4, t8); | |
291 BF(z[3].re, z[1].re, t3, t7); | |
292 BF(z[2].im, z[0].im, t2, t5); | |
293 } | |
294 | |
295 static void fft8(FFTComplex *z) | |
296 { | |
297 FFTSample t1, t2, t3, t4, t5, t6, t7, t8; | |
298 | |
299 fft4(z); | |
300 | |
301 BF(t1, z[5].re, z[4].re, -z[5].re); | |
302 BF(t2, z[5].im, z[4].im, -z[5].im); | |
303 BF(t3, z[7].re, z[6].re, -z[7].re); | |
304 BF(t4, z[7].im, z[6].im, -z[7].im); | |
305 BF(t8, t1, t3, t1); | |
306 BF(t7, t2, t2, t4); | |
307 BF(z[4].re, z[0].re, z[0].re, t1); | |
308 BF(z[4].im, z[0].im, z[0].im, t2); | |
309 BF(z[6].re, z[2].re, z[2].re, t7); | |
310 BF(z[6].im, z[2].im, z[2].im, t8); | |
311 | |
312 TRANSFORM(z[1],z[3],z[5],z[7],sqrthalf,sqrthalf); | |
313 } | |
314 | |
8590 | 315 #if !CONFIG_SMALL |
7542 | 316 static void fft16(FFTComplex *z) |
317 { | |
318 FFTSample t1, t2, t3, t4, t5, t6; | |
319 | |
320 fft8(z); | |
321 fft4(z+8); | |
322 fft4(z+12); | |
323 | |
324 TRANSFORM_ZERO(z[0],z[4],z[8],z[12]); | |
325 TRANSFORM(z[2],z[6],z[10],z[14],sqrthalf,sqrthalf); | |
326 TRANSFORM(z[1],z[5],z[9],z[13],ff_cos_16[1],ff_cos_16[3]); | |
327 TRANSFORM(z[3],z[7],z[11],z[15],ff_cos_16[3],ff_cos_16[1]); | |
328 } | |
329 #else | |
330 DECL_FFT(16,8,4) | |
331 #endif | |
332 DECL_FFT(32,16,8) | |
333 DECL_FFT(64,32,16) | |
334 DECL_FFT(128,64,32) | |
335 DECL_FFT(256,128,64) | |
336 DECL_FFT(512,256,128) | |
8590 | 337 #if !CONFIG_SMALL |
7542 | 338 #define pass pass_big |
339 #endif | |
340 DECL_FFT(1024,512,256) | |
341 DECL_FFT(2048,1024,512) | |
342 DECL_FFT(4096,2048,1024) | |
343 DECL_FFT(8192,4096,2048) | |
344 DECL_FFT(16384,8192,4096) | |
345 DECL_FFT(32768,16384,8192) | |
346 DECL_FFT(65536,32768,16384) | |
347 | |
10391 | 348 static void (* const fft_dispatch[])(FFTComplex*) = { |
7542 | 349 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024, |
350 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536, | |
351 }; | |
352 | |
353 void ff_fft_calc_c(FFTContext *s, FFTComplex *z) | |
354 { | |
355 fft_dispatch[s->nbits-2](z); | |
356 } | |
357 |