00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00027 #include "dsputil.h"
00028
00033 int ff_fft_init(FFTContext *s, int nbits, int inverse)
00034 {
00035 int i, j, m, n;
00036 float alpha, c1, s1, s2;
00037
00038 s->nbits = nbits;
00039 n = 1 << nbits;
00040
00041 s->exptab = av_malloc((n / 2) * sizeof(FFTComplex));
00042 if (!s->exptab)
00043 goto fail;
00044 s->revtab = av_malloc(n * sizeof(uint16_t));
00045 if (!s->revtab)
00046 goto fail;
00047 s->inverse = inverse;
00048
00049 s2 = inverse ? 1.0 : -1.0;
00050
00051 for(i=0;i<(n/2);i++) {
00052 alpha = 2 * M_PI * (float)i / (float)n;
00053 c1 = cos(alpha);
00054 s1 = sin(alpha) * s2;
00055 s->exptab[i].re = c1;
00056 s->exptab[i].im = s1;
00057 }
00058 s->fft_calc = ff_fft_calc_c;
00059 s->imdct_calc = ff_imdct_calc;
00060 s->exptab1 = NULL;
00061
00062
00063 #if defined(HAVE_MMX) \
00064 || (defined(HAVE_ALTIVEC) && !defined(ALTIVEC_USE_REFERENCE_C_CODE))
00065 {
00066 int has_vectors = mm_support();
00067
00068 if (has_vectors) {
00069 #if defined(HAVE_MMX)
00070 if (has_vectors & MM_3DNOWEXT) {
00071
00072 s->imdct_calc = ff_imdct_calc_3dn2;
00073 s->fft_calc = ff_fft_calc_3dn2;
00074 } else if (has_vectors & MM_3DNOW) {
00075
00076 s->fft_calc = ff_fft_calc_3dn;
00077 } else if (has_vectors & MM_SSE) {
00078
00079 s->imdct_calc = ff_imdct_calc_sse;
00080 s->fft_calc = ff_fft_calc_sse;
00081 }
00082 #else
00083 if (has_vectors & MM_ALTIVEC)
00084 s->fft_calc = ff_fft_calc_altivec;
00085 #endif
00086 }
00087 if (s->fft_calc != ff_fft_calc_c) {
00088 int np, nblocks, np2, l;
00089 FFTComplex *q;
00090
00091 np = 1 << nbits;
00092 nblocks = np >> 3;
00093 np2 = np >> 1;
00094 s->exptab1 = av_malloc(np * 2 * sizeof(FFTComplex));
00095 if (!s->exptab1)
00096 goto fail;
00097 q = s->exptab1;
00098 do {
00099 for(l = 0; l < np2; l += 2 * nblocks) {
00100 *q++ = s->exptab[l];
00101 *q++ = s->exptab[l + nblocks];
00102
00103 q->re = -s->exptab[l].im;
00104 q->im = s->exptab[l].re;
00105 q++;
00106 q->re = -s->exptab[l + nblocks].im;
00107 q->im = s->exptab[l + nblocks].re;
00108 q++;
00109 }
00110 nblocks = nblocks >> 1;
00111 } while (nblocks != 0);
00112 av_freep(&s->exptab);
00113 }
00114 }
00115 #endif
00116
00117
00118
00119 for(i=0;i<n;i++) {
00120 m=0;
00121 for(j=0;j<nbits;j++) {
00122 m |= ((i >> j) & 1) << (nbits-j-1);
00123 }
00124 s->revtab[i]=m;
00125 }
00126 return 0;
00127 fail:
00128 av_freep(&s->revtab);
00129 av_freep(&s->exptab);
00130 av_freep(&s->exptab1);
00131 return -1;
00132 }
00133
00134
00135 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
00136 {\
00137 FFTSample ax, ay, bx, by;\
00138 bx=pre1;\
00139 by=pim1;\
00140 ax=qre1;\
00141 ay=qim1;\
00142 pre = (bx + ax);\
00143 pim = (by + ay);\
00144 qre = (bx - ax);\
00145 qim = (by - ay);\
00146 }
00147
00148 #define MUL16(a,b) ((a) * (b))
00149
00150 #define CMUL(pre, pim, are, aim, bre, bim) \
00151 {\
00152 pre = (MUL16(are, bre) - MUL16(aim, bim));\
00153 pim = (MUL16(are, bim) + MUL16(bre, aim));\
00154 }
00155
00161 void ff_fft_calc_c(FFTContext *s, FFTComplex *z)
00162 {
00163 int ln = s->nbits;
00164 int j, np, np2;
00165 int nblocks, nloops;
00166 register FFTComplex *p, *q;
00167 FFTComplex *exptab = s->exptab;
00168 int l;
00169 FFTSample tmp_re, tmp_im;
00170
00171 np = 1 << ln;
00172
00173
00174
00175 p=&z[0];
00176 j=(np >> 1);
00177 do {
00178 BF(p[0].re, p[0].im, p[1].re, p[1].im,
00179 p[0].re, p[0].im, p[1].re, p[1].im);
00180 p+=2;
00181 } while (--j != 0);
00182
00183
00184
00185
00186 p=&z[0];
00187 j=np >> 2;
00188 if (s->inverse) {
00189 do {
00190 BF(p[0].re, p[0].im, p[2].re, p[2].im,
00191 p[0].re, p[0].im, p[2].re, p[2].im);
00192 BF(p[1].re, p[1].im, p[3].re, p[3].im,
00193 p[1].re, p[1].im, -p[3].im, p[3].re);
00194 p+=4;
00195 } while (--j != 0);
00196 } else {
00197 do {
00198 BF(p[0].re, p[0].im, p[2].re, p[2].im,
00199 p[0].re, p[0].im, p[2].re, p[2].im);
00200 BF(p[1].re, p[1].im, p[3].re, p[3].im,
00201 p[1].re, p[1].im, p[3].im, -p[3].re);
00202 p+=4;
00203 } while (--j != 0);
00204 }
00205
00206
00207 nblocks = np >> 3;
00208 nloops = 1 << 2;
00209 np2 = np >> 1;
00210 do {
00211 p = z;
00212 q = z + nloops;
00213 for (j = 0; j < nblocks; ++j) {
00214 BF(p->re, p->im, q->re, q->im,
00215 p->re, p->im, q->re, q->im);
00216
00217 p++;
00218 q++;
00219 for(l = nblocks; l < np2; l += nblocks) {
00220 CMUL(tmp_re, tmp_im, exptab[l].re, exptab[l].im, q->re, q->im);
00221 BF(p->re, p->im, q->re, q->im,
00222 p->re, p->im, tmp_re, tmp_im);
00223 p++;
00224 q++;
00225 }
00226
00227 p += nloops;
00228 q += nloops;
00229 }
00230 nblocks = nblocks >> 1;
00231 nloops = nloops << 1;
00232 } while (nblocks != 0);
00233 }
00234
00238 void ff_fft_permute(FFTContext *s, FFTComplex *z)
00239 {
00240 int j, k, np;
00241 FFTComplex tmp;
00242 const uint16_t *revtab = s->revtab;
00243
00244
00245 np = 1 << s->nbits;
00246 for(j=0;j<np;j++) {
00247 k = revtab[j];
00248 if (k < j) {
00249 tmp = z[k];
00250 z[k] = z[j];
00251 z[j] = tmp;
00252 }
00253 }
00254 }
00255
00256 void ff_fft_end(FFTContext *s)
00257 {
00258 av_freep(&s->revtab);
00259 av_freep(&s->exptab);
00260 av_freep(&s->exptab1);
00261 }
00262