Mercurial > libavcodec.hg
annotate fft.c @ 3823:41ad23b61a2c libavcodec
Original Commit: r17 | ods15 | 2006-09-22 12:28:28 +0300 (Fri, 22 Sep 2006) | 2 lines
bug fixes in codebook header syntax
author | ods15 |
---|---|
date | Mon, 02 Oct 2006 05:55:43 +0000 |
parents | 2ec498208c6a |
children | c8c591fe26f8 |
rev | line source |
---|---|
781 | 1 /* |
2 * FFT/IFFT transforms | |
3 * Copyright (c) 2002 Fabrice Bellard. | |
4 * | |
5 * This library is free software; you can redistribute it and/or | |
6 * modify it under the terms of the GNU Lesser General Public | |
7 * License as published by the Free Software Foundation; either | |
8 * version 2 of the License, or (at your option) any later version. | |
9 * | |
10 * This library is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 * Lesser General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU Lesser General Public | |
16 * License along with this library; if not, write to the Free Software | |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
781 | 18 */ |
1106 | 19 |
20 /** | |
21 * @file fft.c | |
22 * FFT/IFFT transforms. | |
23 */ | |
24 | |
781 | 25 #include "dsputil.h" |
26 | |
27 /** | |
28 * The size of the FFT is 2^nbits. If inverse is TRUE, inverse FFT is | |
2967 | 29 * done |
781 | 30 */ |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1106
diff
changeset
|
31 int ff_fft_init(FFTContext *s, int nbits, int inverse) |
781 | 32 { |
33 int i, j, m, n; | |
34 float alpha, c1, s1, s2; | |
2967 | 35 |
781 | 36 s->nbits = nbits; |
37 n = 1 << nbits; | |
38 | |
39 s->exptab = av_malloc((n / 2) * sizeof(FFTComplex)); | |
40 if (!s->exptab) | |
41 goto fail; | |
42 s->revtab = av_malloc(n * sizeof(uint16_t)); | |
43 if (!s->revtab) | |
44 goto fail; | |
45 s->inverse = inverse; | |
46 | |
47 s2 = inverse ? 1.0 : -1.0; | |
2967 | 48 |
781 | 49 for(i=0;i<(n/2);i++) { |
50 alpha = 2 * M_PI * (float)i / (float)n; | |
51 c1 = cos(alpha); | |
52 s1 = sin(alpha) * s2; | |
53 s->exptab[i].re = c1; | |
54 s->exptab[i].im = s1; | |
55 } | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1106
diff
changeset
|
56 s->fft_calc = ff_fft_calc_c; |
3555 | 57 s->imdct_calc = ff_imdct_calc; |
781 | 58 s->exptab1 = NULL; |
59 | |
60 /* compute constant table for HAVE_SSE version */ | |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3567
diff
changeset
|
61 #if defined(HAVE_MMX) \ |
3567
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
62 || (defined(HAVE_ALTIVEC) && !defined(ALTIVEC_USE_REFERENCE_C_CODE)) |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
63 { |
3567
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
64 int has_vectors = mm_support(); |
781 | 65 |
3567
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
66 if (has_vectors) { |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
67 #if defined(HAVE_MMX) |
3746 | 68 if (has_vectors & MM_3DNOWEXT) { |
69 /* 3DNowEx for K7/K8 */ | |
3567
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
70 s->imdct_calc = ff_imdct_calc_3dn2; |
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
71 s->fft_calc = ff_fft_calc_3dn2; |
3746 | 72 } else if (has_vectors & MM_3DNOW) { |
3567
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
73 /* 3DNow! for K6-2/3 */ |
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
74 s->fft_calc = ff_fft_calc_3dn; |
3746 | 75 } else if (has_vectors & MM_SSE) { |
3591 | 76 /* SSE for P3/P4 */ |
3746 | 77 s->imdct_calc = ff_imdct_calc_sse; |
3567
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
78 s->fft_calc = ff_fft_calc_sse; |
3746 | 79 } |
3567
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
80 #else /* HAVE_MMX */ |
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
81 if (has_vectors & MM_ALTIVEC) |
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
82 s->fft_calc = ff_fft_calc_altivec; |
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
83 #endif |
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
84 } |
1f8730f62765
r5954 broke fft on cpus with 3dnow but without mm3dnow.h
lorenm
parents:
3555
diff
changeset
|
85 if (s->fft_calc != ff_fft_calc_c) { |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
86 int np, nblocks, np2, l; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
87 FFTComplex *q; |
2967 | 88 |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
89 np = 1 << nbits; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
90 nblocks = np >> 3; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
91 np2 = np >> 1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
92 s->exptab1 = av_malloc(np * 2 * sizeof(FFTComplex)); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
93 if (!s->exptab1) |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
94 goto fail; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
95 q = s->exptab1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
96 do { |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
97 for(l = 0; l < np2; l += 2 * nblocks) { |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
98 *q++ = s->exptab[l]; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
99 *q++ = s->exptab[l + nblocks]; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
100 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
101 q->re = -s->exptab[l].im; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
102 q->im = s->exptab[l].re; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
103 q++; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
104 q->re = -s->exptab[l + nblocks].im; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
105 q->im = s->exptab[l + nblocks].re; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
106 q++; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
107 } |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
108 nblocks = nblocks >> 1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
109 } while (nblocks != 0); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
110 av_freep(&s->exptab); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
111 } |
781 | 112 } |
113 #endif | |
114 | |
115 /* compute bit reverse table */ | |
116 | |
117 for(i=0;i<n;i++) { | |
118 m=0; | |
119 for(j=0;j<nbits;j++) { | |
120 m |= ((i >> j) & 1) << (nbits-j-1); | |
121 } | |
122 s->revtab[i]=m; | |
123 } | |
124 return 0; | |
125 fail: | |
126 av_freep(&s->revtab); | |
127 av_freep(&s->exptab); | |
128 av_freep(&s->exptab1); | |
129 return -1; | |
130 } | |
131 | |
132 /* butter fly op */ | |
133 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \ | |
134 {\ | |
135 FFTSample ax, ay, bx, by;\ | |
136 bx=pre1;\ | |
137 by=pim1;\ | |
138 ax=qre1;\ | |
139 ay=qim1;\ | |
140 pre = (bx + ax);\ | |
141 pim = (by + ay);\ | |
142 qre = (bx - ax);\ | |
143 qim = (by - ay);\ | |
144 } | |
145 | |
146 #define MUL16(a,b) ((a) * (b)) | |
147 | |
148 #define CMUL(pre, pim, are, aim, bre, bim) \ | |
149 {\ | |
150 pre = (MUL16(are, bre) - MUL16(aim, bim));\ | |
151 pim = (MUL16(are, bim) + MUL16(bre, aim));\ | |
152 } | |
153 | |
154 /** | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1106
diff
changeset
|
155 * Do a complex FFT with the parameters defined in ff_fft_init(). The |
781 | 156 * input data must be permuted before with s->revtab table. No |
2967 | 157 * 1.0/sqrt(n) normalization is done. |
781 | 158 */ |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1106
diff
changeset
|
159 void ff_fft_calc_c(FFTContext *s, FFTComplex *z) |
781 | 160 { |
161 int ln = s->nbits; | |
2979 | 162 int j, np, np2; |
163 int nblocks, nloops; | |
781 | 164 register FFTComplex *p, *q; |
165 FFTComplex *exptab = s->exptab; | |
166 int l; | |
167 FFTSample tmp_re, tmp_im; | |
168 | |
169 np = 1 << ln; | |
170 | |
171 /* pass 0 */ | |
172 | |
173 p=&z[0]; | |
174 j=(np >> 1); | |
175 do { | |
2967 | 176 BF(p[0].re, p[0].im, p[1].re, p[1].im, |
781 | 177 p[0].re, p[0].im, p[1].re, p[1].im); |
178 p+=2; | |
179 } while (--j != 0); | |
180 | |
181 /* pass 1 */ | |
182 | |
2967 | 183 |
781 | 184 p=&z[0]; |
185 j=np >> 2; | |
186 if (s->inverse) { | |
187 do { | |
2967 | 188 BF(p[0].re, p[0].im, p[2].re, p[2].im, |
781 | 189 p[0].re, p[0].im, p[2].re, p[2].im); |
2967 | 190 BF(p[1].re, p[1].im, p[3].re, p[3].im, |
781 | 191 p[1].re, p[1].im, -p[3].im, p[3].re); |
192 p+=4; | |
193 } while (--j != 0); | |
194 } else { | |
195 do { | |
2967 | 196 BF(p[0].re, p[0].im, p[2].re, p[2].im, |
781 | 197 p[0].re, p[0].im, p[2].re, p[2].im); |
2967 | 198 BF(p[1].re, p[1].im, p[3].re, p[3].im, |
781 | 199 p[1].re, p[1].im, p[3].im, -p[3].re); |
200 p+=4; | |
201 } while (--j != 0); | |
202 } | |
203 /* pass 2 .. ln-1 */ | |
204 | |
205 nblocks = np >> 3; | |
206 nloops = 1 << 2; | |
207 np2 = np >> 1; | |
208 do { | |
209 p = z; | |
210 q = z + nloops; | |
211 for (j = 0; j < nblocks; ++j) { | |
212 BF(p->re, p->im, q->re, q->im, | |
213 p->re, p->im, q->re, q->im); | |
2967 | 214 |
781 | 215 p++; |
216 q++; | |
217 for(l = nblocks; l < np2; l += nblocks) { | |
218 CMUL(tmp_re, tmp_im, exptab[l].re, exptab[l].im, q->re, q->im); | |
219 BF(p->re, p->im, q->re, q->im, | |
220 p->re, p->im, tmp_re, tmp_im); | |
221 p++; | |
222 q++; | |
223 } | |
224 | |
225 p += nloops; | |
226 q += nloops; | |
227 } | |
228 nblocks = nblocks >> 1; | |
229 nloops = nloops << 1; | |
230 } while (nblocks != 0); | |
231 } | |
232 | |
233 /** | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1106
diff
changeset
|
234 * Do the permutation needed BEFORE calling ff_fft_calc() |
781 | 235 */ |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1106
diff
changeset
|
236 void ff_fft_permute(FFTContext *s, FFTComplex *z) |
781 | 237 { |
238 int j, k, np; | |
239 FFTComplex tmp; | |
240 const uint16_t *revtab = s->revtab; | |
2967 | 241 |
781 | 242 /* reverse */ |
243 np = 1 << s->nbits; | |
244 for(j=0;j<np;j++) { | |
245 k = revtab[j]; | |
246 if (k < j) { | |
247 tmp = z[k]; | |
248 z[k] = z[j]; | |
249 z[j] = tmp; | |
250 } | |
251 } | |
252 } | |
253 | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1106
diff
changeset
|
254 void ff_fft_end(FFTContext *s) |
781 | 255 { |
256 av_freep(&s->revtab); | |
257 av_freep(&s->exptab); | |
258 av_freep(&s->exptab1); | |
259 } | |
260 |