enabled SSE fft on x86
[libav.git] / libavcodec / fft.c
CommitLineData
bb6f5690
FB
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
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19#include "dsputil.h"
20
21/**
22 * The size of the FFT is 2^nbits. If inverse is TRUE, inverse FFT is
23 * done
24 */
25int fft_init(FFTContext *s, int nbits, int inverse)
26{
27 int i, j, m, n;
28 float alpha, c1, s1, s2;
29
30 s->nbits = nbits;
31 n = 1 << nbits;
32
33 s->exptab = av_malloc((n / 2) * sizeof(FFTComplex));
34 if (!s->exptab)
35 goto fail;
36 s->revtab = av_malloc(n * sizeof(uint16_t));
37 if (!s->revtab)
38 goto fail;
39 s->inverse = inverse;
40
41 s2 = inverse ? 1.0 : -1.0;
42
43 for(i=0;i<(n/2);i++) {
44 alpha = 2 * M_PI * (float)i / (float)n;
45 c1 = cos(alpha);
46 s1 = sin(alpha) * s2;
47 s->exptab[i].re = c1;
48 s->exptab[i].im = s1;
49 }
50 s->fft_calc = fft_calc_c;
51 s->exptab1 = NULL;
52
53 /* compute constant table for HAVE_SSE version */
bbbb6d6f
FB
54#if defined(HAVE_MMX) && defined(HAVE_BUILTIN_VECTOR)
55 if (mm_support() & MM_SSE) {
bb6f5690
FB
56 int np, nblocks, np2, l;
57 FFTComplex *q;
bbbb6d6f 58
bb6f5690
FB
59 np = 1 << nbits;
60 nblocks = np >> 3;
61 np2 = np >> 1;
62 s->exptab1 = av_malloc(np * 2 * sizeof(FFTComplex));
63 if (!s->exptab1)
64 goto fail;
65 q = s->exptab1;
66 do {
67 for(l = 0; l < np2; l += 2 * nblocks) {
68 *q++ = s->exptab[l];
69 *q++ = s->exptab[l + nblocks];
70
71 q->re = -s->exptab[l].im;
72 q->im = s->exptab[l].re;
73 q++;
74 q->re = -s->exptab[l + nblocks].im;
75 q->im = s->exptab[l + nblocks].re;
76 q++;
77 }
78 nblocks = nblocks >> 1;
79 } while (nblocks != 0);
80 av_freep(&s->exptab);
bbbb6d6f 81 s->fft_calc = fft_calc_sse;
bb6f5690
FB
82 }
83#endif
84
85 /* compute bit reverse table */
86
87 for(i=0;i<n;i++) {
88 m=0;
89 for(j=0;j<nbits;j++) {
90 m |= ((i >> j) & 1) << (nbits-j-1);
91 }
92 s->revtab[i]=m;
93 }
94 return 0;
95 fail:
96 av_freep(&s->revtab);
97 av_freep(&s->exptab);
98 av_freep(&s->exptab1);
99 return -1;
100}
101
102/* butter fly op */
103#define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
104{\
105 FFTSample ax, ay, bx, by;\
106 bx=pre1;\
107 by=pim1;\
108 ax=qre1;\
109 ay=qim1;\
110 pre = (bx + ax);\
111 pim = (by + ay);\
112 qre = (bx - ax);\
113 qim = (by - ay);\
114}
115
116#define MUL16(a,b) ((a) * (b))
117
118#define CMUL(pre, pim, are, aim, bre, bim) \
119{\
120 pre = (MUL16(are, bre) - MUL16(aim, bim));\
121 pim = (MUL16(are, bim) + MUL16(bre, aim));\
122}
123
124/**
125 * Do a complex FFT with the parameters defined in fft_init(). The
126 * input data must be permuted before with s->revtab table. No
127 * 1.0/sqrt(n) normalization is done.
128 */
129void fft_calc_c(FFTContext *s, FFTComplex *z)
130{
131 int ln = s->nbits;
132 int j, np, np2;
133 int nblocks, nloops;
134 register FFTComplex *p, *q;
135 FFTComplex *exptab = s->exptab;
136 int l;
137 FFTSample tmp_re, tmp_im;
138
139 np = 1 << ln;
140
141 /* pass 0 */
142
143 p=&z[0];
144 j=(np >> 1);
145 do {
146 BF(p[0].re, p[0].im, p[1].re, p[1].im,
147 p[0].re, p[0].im, p[1].re, p[1].im);
148 p+=2;
149 } while (--j != 0);
150
151 /* pass 1 */
152
153
154 p=&z[0];
155 j=np >> 2;
156 if (s->inverse) {
157 do {
158 BF(p[0].re, p[0].im, p[2].re, p[2].im,
159 p[0].re, p[0].im, p[2].re, p[2].im);
160 BF(p[1].re, p[1].im, p[3].re, p[3].im,
161 p[1].re, p[1].im, -p[3].im, p[3].re);
162 p+=4;
163 } while (--j != 0);
164 } else {
165 do {
166 BF(p[0].re, p[0].im, p[2].re, p[2].im,
167 p[0].re, p[0].im, p[2].re, p[2].im);
168 BF(p[1].re, p[1].im, p[3].re, p[3].im,
169 p[1].re, p[1].im, p[3].im, -p[3].re);
170 p+=4;
171 } while (--j != 0);
172 }
173 /* pass 2 .. ln-1 */
174
175 nblocks = np >> 3;
176 nloops = 1 << 2;
177 np2 = np >> 1;
178 do {
179 p = z;
180 q = z + nloops;
181 for (j = 0; j < nblocks; ++j) {
182 BF(p->re, p->im, q->re, q->im,
183 p->re, p->im, q->re, q->im);
184
185 p++;
186 q++;
187 for(l = nblocks; l < np2; l += nblocks) {
188 CMUL(tmp_re, tmp_im, exptab[l].re, exptab[l].im, q->re, q->im);
189 BF(p->re, p->im, q->re, q->im,
190 p->re, p->im, tmp_re, tmp_im);
191 p++;
192 q++;
193 }
194
195 p += nloops;
196 q += nloops;
197 }
198 nblocks = nblocks >> 1;
199 nloops = nloops << 1;
200 } while (nblocks != 0);
201}
202
203/**
204 * Do the permutation needed BEFORE calling fft_calc()
205 */
206void fft_permute(FFTContext *s, FFTComplex *z)
207{
208 int j, k, np;
209 FFTComplex tmp;
210 const uint16_t *revtab = s->revtab;
211
212 /* reverse */
213 np = 1 << s->nbits;
214 for(j=0;j<np;j++) {
215 k = revtab[j];
216 if (k < j) {
217 tmp = z[k];
218 z[k] = z[j];
219 z[j] = tmp;
220 }
221 }
222}
223
224void fft_end(FFTContext *s)
225{
226 av_freep(&s->revtab);
227 av_freep(&s->exptab);
228 av_freep(&s->exptab1);
229}
230