annotate fft.c @ 8673:5b7d5a33a3e2 libavcodec

Fix filenames in Doxygen comments.
author diego
date Mon, 26 Jan 2009 22:51:16 +0000
parents 3f72756b0c5c
children 04200b454f3f
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
1 /*
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
2 * FFT/IFFT transforms
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
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
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
5 * Partly based on libdjbfft by D. J. Bernstein
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
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
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
10 * modify it under the terms of the GNU Lesser General Public
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
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
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
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
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
17 * Lesser General Public License for more details.
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
18 *
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
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
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
22 */
1106
1e39f273ecd6 per file doxy
michaelni
parents: 1009
diff changeset
23
1e39f273ecd6 per file doxy
michaelni
parents: 1009
diff changeset
24 /**
1e39f273ecd6 per file doxy
michaelni
parents: 1009
diff changeset
25 * @file fft.c
1e39f273ecd6 per file doxy
michaelni
parents: 1009
diff changeset
26 * FFT/IFFT transforms.
1e39f273ecd6 per file doxy
michaelni
parents: 1009
diff changeset
27 */
1e39f273ecd6 per file doxy
michaelni
parents: 1009
diff changeset
28
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
29 #include "dsputil.h"
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
30
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
31 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
32 DECLARE_ALIGNED_16(FFTSample, ff_cos_16[8]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
33 DECLARE_ALIGNED_16(FFTSample, ff_cos_32[16]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
34 DECLARE_ALIGNED_16(FFTSample, ff_cos_64[32]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
35 DECLARE_ALIGNED_16(FFTSample, ff_cos_128[64]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
36 DECLARE_ALIGNED_16(FFTSample, ff_cos_256[128]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
37 DECLARE_ALIGNED_16(FFTSample, ff_cos_512[256]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
38 DECLARE_ALIGNED_16(FFTSample, ff_cos_1024[512]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
39 DECLARE_ALIGNED_16(FFTSample, ff_cos_2048[1024]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
40 DECLARE_ALIGNED_16(FFTSample, ff_cos_4096[2048]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
41 DECLARE_ALIGNED_16(FFTSample, ff_cos_8192[4096]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
42 DECLARE_ALIGNED_16(FFTSample, ff_cos_16384[8192]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
43 DECLARE_ALIGNED_16(FFTSample, ff_cos_32768[16384]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
44 DECLARE_ALIGNED_16(FFTSample, ff_cos_65536[32768]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
45 static FFTSample *ff_cos_tabs[] = {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
46 ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256, ff_cos_512, ff_cos_1024,
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
47 ff_cos_2048, ff_cos_4096, ff_cos_8192, ff_cos_16384, ff_cos_32768, ff_cos_65536,
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
48 };
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
49
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
50 static int split_radix_permutation(int i, int n, int inverse)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
51 {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
52 int m;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
53 if(n <= 2) return i&1;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
54 m = n >> 1;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
55 if(!(i&m)) return split_radix_permutation(i, m, inverse)*2;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
56 m >>= 1;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
57 if(inverse == !(i&m)) return split_radix_permutation(i, m, inverse)*4 + 1;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
58 else return split_radix_permutation(i, m, inverse)*4 - 1;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
59 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
60
8637
3f72756b0c5c Mark ff_fft_init with av_cold.
alexc
parents: 8636
diff changeset
61 av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
62 {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
63 int i, j, m, n;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
64 float alpha, c1, s1, s2;
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
65 int split_radix = 1;
6503
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
66 int av_unused has_vectors;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1879
diff changeset
67
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
68 if (nbits < 2 || nbits > 16)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
69 goto fail;
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
70 s->nbits = nbits;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
71 n = 1 << nbits;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
72
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
73 s->tmp_buf = NULL;
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
74 s->exptab = av_malloc((n / 2) * sizeof(FFTComplex));
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
75 if (!s->exptab)
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
76 goto fail;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
77 s->revtab = av_malloc(n * sizeof(uint16_t));
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
78 if (!s->revtab)
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
79 goto fail;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
80 s->inverse = inverse;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
81
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
82 s2 = inverse ? 1.0 : -1.0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1879
diff changeset
83
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
84 s->fft_permute = ff_fft_permute_c;
1879
dd63cb7e5080 fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents: 1106
diff changeset
85 s->fft_calc = ff_fft_calc_c;
7547
8226017a65ae mdct wrapper function to match fft
lorenm
parents: 7544
diff changeset
86 s->imdct_calc = ff_imdct_calc_c;
8226017a65ae mdct wrapper function to match fft
lorenm
parents: 7544
diff changeset
87 s->imdct_half = ff_imdct_half_c;
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
88 s->exptab1 = NULL;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
89
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8104
diff changeset
90 #if HAVE_MMX && HAVE_YASM
6503
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
91 has_vectors = mm_support();
8104
0d108ec85620 Remove duplicated MM_* macros for CPU capabilities from dsputil.h.
rathann
parents: 7547
diff changeset
92 if (has_vectors & FF_MM_SSE) {
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
93 /* SSE for P3/P4/K8 */
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
94 s->imdct_calc = ff_imdct_calc_sse;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
95 s->imdct_half = ff_imdct_half_sse;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
96 s->fft_permute = ff_fft_permute_sse;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
97 s->fft_calc = ff_fft_calc_sse;
8104
0d108ec85620 Remove duplicated MM_* macros for CPU capabilities from dsputil.h.
rathann
parents: 7547
diff changeset
98 } else if (has_vectors & FF_MM_3DNOWEXT) {
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
99 /* 3DNowEx for K7 */
6503
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
100 s->imdct_calc = ff_imdct_calc_3dn2;
7263
fc843d00867c exploit mdct symmetry
lorenm
parents: 6504
diff changeset
101 s->imdct_half = ff_imdct_half_3dn2;
6503
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
102 s->fft_calc = ff_fft_calc_3dn2;
8104
0d108ec85620 Remove duplicated MM_* macros for CPU capabilities from dsputil.h.
rathann
parents: 7547
diff changeset
103 } else if (has_vectors & FF_MM_3DNOW) {
6503
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
104 /* 3DNow! for K6-2/3 */
7544
ee1cb5ab9f99 optimize imdct_half:
lorenm
parents: 7543
diff changeset
105 s->imdct_calc = ff_imdct_calc_3dn;
ee1cb5ab9f99 optimize imdct_half:
lorenm
parents: 7543
diff changeset
106 s->imdct_half = ff_imdct_half_3dn;
6503
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
107 s->fft_calc = ff_fft_calc_3dn;
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
108 }
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8104
diff changeset
109 #elif HAVE_ALTIVEC && !defined ALTIVEC_USE_REFERENCE_C_CODE
6503
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
110 has_vectors = mm_support();
8104
0d108ec85620 Remove duplicated MM_* macros for CPU capabilities from dsputil.h.
rathann
parents: 7547
diff changeset
111 if (has_vectors & FF_MM_ALTIVEC) {
6503
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
112 s->fft_calc = ff_fft_calc_altivec;
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
113 split_radix = 0;
6503
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
114 }
311c95dddb16 clean up FFT SIMD selection
mru
parents: 3947
diff changeset
115 #endif
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
116
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
117 if (split_radix) {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
118 for(j=4; j<=nbits; j++) {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
119 int m = 1<<j;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
120 double freq = 2*M_PI/m;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
121 FFTSample *tab = ff_cos_tabs[j-4];
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
122 for(i=0; i<=m/4; i++)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
123 tab[i] = cos(i*freq);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
124 for(i=1; i<m/4; i++)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
125 tab[m/2-i] = tab[i];
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
126 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
127 for(i=0; i<n; i++)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
128 s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = i;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
129 s->tmp_buf = av_malloc(n * sizeof(FFTComplex));
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
130 } else {
6504
978911ca654f indentation
mru
parents: 6503
diff changeset
131 int np, nblocks, np2, l;
978911ca654f indentation
mru
parents: 6503
diff changeset
132 FFTComplex *q;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1879
diff changeset
133
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
134 for(i=0; i<(n/2); i++) {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
135 alpha = 2 * M_PI * (float)i / (float)n;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
136 c1 = cos(alpha);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
137 s1 = sin(alpha) * s2;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
138 s->exptab[i].re = c1;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
139 s->exptab[i].im = s1;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
140 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
141
6504
978911ca654f indentation
mru
parents: 6503
diff changeset
142 np = 1 << nbits;
978911ca654f indentation
mru
parents: 6503
diff changeset
143 nblocks = np >> 3;
978911ca654f indentation
mru
parents: 6503
diff changeset
144 np2 = np >> 1;
978911ca654f indentation
mru
parents: 6503
diff changeset
145 s->exptab1 = av_malloc(np * 2 * sizeof(FFTComplex));
978911ca654f indentation
mru
parents: 6503
diff changeset
146 if (!s->exptab1)
978911ca654f indentation
mru
parents: 6503
diff changeset
147 goto fail;
978911ca654f indentation
mru
parents: 6503
diff changeset
148 q = s->exptab1;
978911ca654f indentation
mru
parents: 6503
diff changeset
149 do {
978911ca654f indentation
mru
parents: 6503
diff changeset
150 for(l = 0; l < np2; l += 2 * nblocks) {
978911ca654f indentation
mru
parents: 6503
diff changeset
151 *q++ = s->exptab[l];
978911ca654f indentation
mru
parents: 6503
diff changeset
152 *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
153
6504
978911ca654f indentation
mru
parents: 6503
diff changeset
154 q->re = -s->exptab[l].im;
978911ca654f indentation
mru
parents: 6503
diff changeset
155 q->im = s->exptab[l].re;
978911ca654f indentation
mru
parents: 6503
diff changeset
156 q++;
978911ca654f indentation
mru
parents: 6503
diff changeset
157 q->re = -s->exptab[l + nblocks].im;
978911ca654f indentation
mru
parents: 6503
diff changeset
158 q->im = s->exptab[l + nblocks].re;
978911ca654f indentation
mru
parents: 6503
diff changeset
159 q++;
978911ca654f indentation
mru
parents: 6503
diff changeset
160 }
978911ca654f indentation
mru
parents: 6503
diff changeset
161 nblocks = nblocks >> 1;
978911ca654f indentation
mru
parents: 6503
diff changeset
162 } while (nblocks != 0);
978911ca654f indentation
mru
parents: 6503
diff changeset
163 av_freep(&s->exptab);
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
164
7543
lorenm
parents: 7542
diff changeset
165 /* compute bit reverse table */
lorenm
parents: 7542
diff changeset
166 for(i=0;i<n;i++) {
lorenm
parents: 7542
diff changeset
167 m=0;
lorenm
parents: 7542
diff changeset
168 for(j=0;j<nbits;j++) {
lorenm
parents: 7542
diff changeset
169 m |= ((i >> j) & 1) << (nbits-j-1);
lorenm
parents: 7542
diff changeset
170 }
lorenm
parents: 7542
diff changeset
171 s->revtab[i]=m;
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
172 }
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
173 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
174
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
175 return 0;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
176 fail:
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
177 av_freep(&s->revtab);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
178 av_freep(&s->exptab);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
179 av_freep(&s->exptab1);
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
180 av_freep(&s->tmp_buf);
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
181 return -1;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
182 }
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
183
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
184 void ff_fft_permute_c(FFTContext *s, FFTComplex *z)
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
185 {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
186 int j, k, np;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
187 FFTComplex tmp;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
188 const uint16_t *revtab = s->revtab;
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
189 np = 1 << s->nbits;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
190
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
191 if (s->tmp_buf) {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
192 /* TODO: handle split-radix permute in a more optimal way, probably in-place */
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
193 for(j=0;j<np;j++) s->tmp_buf[revtab[j]] = z[j];
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
194 memcpy(z, s->tmp_buf, np * sizeof(FFTComplex));
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
195 return;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
196 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1879
diff changeset
197
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
198 /* reverse */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
199 for(j=0;j<np;j++) {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
200 k = revtab[j];
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
201 if (k < j) {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
202 tmp = z[k];
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
203 z[k] = z[j];
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
204 z[j] = tmp;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
205 }
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
206 }
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
207 }
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
208
1879
dd63cb7e5080 fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents: 1106
diff changeset
209 void ff_fft_end(FFTContext *s)
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
210 {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
211 av_freep(&s->revtab);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
212 av_freep(&s->exptab);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
213 av_freep(&s->exptab1);
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
214 av_freep(&s->tmp_buf);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
215 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
216
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
217 #define sqrthalf (float)M_SQRT1_2
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
218
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
219 #define BF(x,y,a,b) {\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
220 x = a - b;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
221 y = a + b;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
222 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
223
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
224 #define BUTTERFLIES(a0,a1,a2,a3) {\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
225 BF(t3, t5, t5, t1);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
226 BF(a2.re, a0.re, a0.re, t5);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
227 BF(a3.im, a1.im, a1.im, t3);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
228 BF(t4, t6, t2, t6);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
229 BF(a3.re, a1.re, a1.re, t4);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
230 BF(a2.im, a0.im, a0.im, t6);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
231 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
232
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
233 // force loading all the inputs before storing any.
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
234 // this is slightly slower for small data, but avoids store->load aliasing
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
235 // for addresses separated by large powers of 2.
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
236 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
237 FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
238 BF(t3, t5, t5, t1);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
239 BF(a2.re, a0.re, r0, t5);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
240 BF(a3.im, a1.im, i1, t3);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
241 BF(t4, t6, t2, t6);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
242 BF(a3.re, a1.re, r1, t4);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
243 BF(a2.im, a0.im, i0, t6);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
244 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
245
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
246 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
247 t1 = a2.re * wre + a2.im * wim;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
248 t2 = a2.im * wre - a2.re * wim;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
249 t5 = a3.re * wre - a3.im * wim;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
250 t6 = a3.im * wre + a3.re * wim;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
251 BUTTERFLIES(a0,a1,a2,a3)\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
252 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
253
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
254 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
255 t1 = a2.re;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
256 t2 = a2.im;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
257 t5 = a3.re;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
258 t6 = a3.im;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
259 BUTTERFLIES(a0,a1,a2,a3)\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
260 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
261
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
262 /* z[0...8n-1], w[1...2n-1] */
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
263 #define PASS(name)\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
264 static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
265 {\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
266 FFTSample t1, t2, t3, t4, t5, t6;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
267 int o1 = 2*n;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
268 int o2 = 4*n;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
269 int o3 = 6*n;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
270 const FFTSample *wim = wre+o1;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
271 n--;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
272 \
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
273 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
274 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
275 do {\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
276 z += 2;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
277 wre += 2;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
278 wim -= 2;\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
279 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
280 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
281 } while(--n);\
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
282 }
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents:
diff changeset
283
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
284 PASS(pass)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
285 #undef BUTTERFLIES
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
286 #define BUTTERFLIES BUTTERFLIES_BIG
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
287 PASS(pass_big)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
288
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
289 #define DECL_FFT(n,n2,n4)\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
290 static void fft##n(FFTComplex *z)\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
291 {\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
292 fft##n2(z);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
293 fft##n4(z+n4*2);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
294 fft##n4(z+n4*3);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
295 pass(z,ff_cos_##n,n4/2);\
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
296 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
297
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
298 static void fft4(FFTComplex *z)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
299 {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
300 FFTSample t1, t2, t3, t4, t5, t6, t7, t8;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
301
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
302 BF(t3, t1, z[0].re, z[1].re);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
303 BF(t8, t6, z[3].re, z[2].re);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
304 BF(z[2].re, z[0].re, t1, t6);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
305 BF(t4, t2, z[0].im, z[1].im);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
306 BF(t7, t5, z[2].im, z[3].im);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
307 BF(z[3].im, z[1].im, t4, t8);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
308 BF(z[3].re, z[1].re, t3, t7);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
309 BF(z[2].im, z[0].im, t2, t5);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
310 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
311
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
312 static void fft8(FFTComplex *z)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
313 {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
314 FFTSample t1, t2, t3, t4, t5, t6, t7, t8;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
315
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
316 fft4(z);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
317
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
318 BF(t1, z[5].re, z[4].re, -z[5].re);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
319 BF(t2, z[5].im, z[4].im, -z[5].im);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
320 BF(t3, z[7].re, z[6].re, -z[7].re);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
321 BF(t4, z[7].im, z[6].im, -z[7].im);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
322 BF(t8, t1, t3, t1);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
323 BF(t7, t2, t2, t4);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
324 BF(z[4].re, z[0].re, z[0].re, t1);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
325 BF(z[4].im, z[0].im, z[0].im, t2);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
326 BF(z[6].re, z[2].re, z[2].re, t7);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
327 BF(z[6].im, z[2].im, z[2].im, t8);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
328
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
329 TRANSFORM(z[1],z[3],z[5],z[7],sqrthalf,sqrthalf);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
330 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
331
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8104
diff changeset
332 #if !CONFIG_SMALL
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
333 static void fft16(FFTComplex *z)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
334 {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
335 FFTSample t1, t2, t3, t4, t5, t6;
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
336
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
337 fft8(z);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
338 fft4(z+8);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
339 fft4(z+12);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
340
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
341 TRANSFORM_ZERO(z[0],z[4],z[8],z[12]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
342 TRANSFORM(z[2],z[6],z[10],z[14],sqrthalf,sqrthalf);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
343 TRANSFORM(z[1],z[5],z[9],z[13],ff_cos_16[1],ff_cos_16[3]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
344 TRANSFORM(z[3],z[7],z[11],z[15],ff_cos_16[3],ff_cos_16[1]);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
345 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
346 #else
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
347 DECL_FFT(16,8,4)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
348 #endif
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
349 DECL_FFT(32,16,8)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
350 DECL_FFT(64,32,16)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
351 DECL_FFT(128,64,32)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
352 DECL_FFT(256,128,64)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
353 DECL_FFT(512,256,128)
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8104
diff changeset
354 #if !CONFIG_SMALL
7542
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
355 #define pass pass_big
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
356 #endif
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
357 DECL_FFT(1024,512,256)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
358 DECL_FFT(2048,1024,512)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
359 DECL_FFT(4096,2048,1024)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
360 DECL_FFT(8192,4096,2048)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
361 DECL_FFT(16384,8192,4096)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
362 DECL_FFT(32768,16384,8192)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
363 DECL_FFT(65536,32768,16384)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
364
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
365 static void (*fft_dispatch[])(FFTComplex*) = {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
366 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
367 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536,
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
368 };
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
369
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
370 void ff_fft_calc_c(FFTContext *s, FFTComplex *z)
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
371 {
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
372 fft_dispatch[s->nbits-2](z);
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
373 }
a8a8205a9081 split-radix FFT
lorenm
parents: 7263
diff changeset
374