libavcodec/fft-test.c
Go to the documentation of this file.
00001 /*
00002  * (c) 2002 Fabrice Bellard
00003  *
00004  * This file is part of FFmpeg.
00005  *
00006  * FFmpeg is free software; you can redistribute it and/or
00007  * modify it under the terms of the GNU Lesser General Public
00008  * License as published by the Free Software Foundation; either
00009  * version 2.1 of the License, or (at your option) any later version.
00010  *
00011  * FFmpeg is distributed in the hope that it will be useful,
00012  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00013  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00014  * Lesser General Public License for more details.
00015  *
00016  * You should have received a copy of the GNU Lesser General Public
00017  * License along with FFmpeg; if not, write to the Free Software
00018  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00019  */
00020 
00026 #include "libavutil/mathematics.h"
00027 #include "libavutil/lfg.h"
00028 #include "libavutil/log.h"
00029 #include "fft.h"
00030 #if CONFIG_FFT_FLOAT
00031 #include "dct.h"
00032 #include "rdft.h"
00033 #endif
00034 #include <math.h>
00035 #include <unistd.h>
00036 #include <sys/time.h>
00037 #include <stdlib.h>
00038 #include <string.h>
00039 
00040 /* reference fft */
00041 
00042 #define MUL16(a,b) ((a) * (b))
00043 
00044 #define CMAC(pre, pim, are, aim, bre, bim) \
00045 {\
00046    pre += (MUL16(are, bre) - MUL16(aim, bim));\
00047    pim += (MUL16(are, bim) + MUL16(bre, aim));\
00048 }
00049 
00050 #if CONFIG_FFT_FLOAT
00051 #   define RANGE 1.0
00052 #   define REF_SCALE(x, bits)  (x)
00053 #   define FMT "%10.6f"
00054 #else
00055 #   define RANGE 16384
00056 #   define REF_SCALE(x, bits) ((x) / (1<<(bits)))
00057 #   define FMT "%6d"
00058 #endif
00059 
00060 struct {
00061     float re, im;
00062 } *exptab;
00063 
00064 static void fft_ref_init(int nbits, int inverse)
00065 {
00066     int n, i;
00067     double c1, s1, alpha;
00068 
00069     n = 1 << nbits;
00070     exptab = av_malloc((n / 2) * sizeof(*exptab));
00071 
00072     for (i = 0; i < (n/2); i++) {
00073         alpha = 2 * M_PI * (float)i / (float)n;
00074         c1 = cos(alpha);
00075         s1 = sin(alpha);
00076         if (!inverse)
00077             s1 = -s1;
00078         exptab[i].re = c1;
00079         exptab[i].im = s1;
00080     }
00081 }
00082 
00083 static void fft_ref(FFTComplex *tabr, FFTComplex *tab, int nbits)
00084 {
00085     int n, i, j, k, n2;
00086     double tmp_re, tmp_im, s, c;
00087     FFTComplex *q;
00088 
00089     n = 1 << nbits;
00090     n2 = n >> 1;
00091     for (i = 0; i < n; i++) {
00092         tmp_re = 0;
00093         tmp_im = 0;
00094         q = tab;
00095         for (j = 0; j < n; j++) {
00096             k = (i * j) & (n - 1);
00097             if (k >= n2) {
00098                 c = -exptab[k - n2].re;
00099                 s = -exptab[k - n2].im;
00100             } else {
00101                 c = exptab[k].re;
00102                 s = exptab[k].im;
00103             }
00104             CMAC(tmp_re, tmp_im, c, s, q->re, q->im);
00105             q++;
00106         }
00107         tabr[i].re = REF_SCALE(tmp_re, nbits);
00108         tabr[i].im = REF_SCALE(tmp_im, nbits);
00109     }
00110 }
00111 
00112 static void imdct_ref(FFTSample *out, FFTSample *in, int nbits)
00113 {
00114     int n = 1<<nbits;
00115     int k, i, a;
00116     double sum, f;
00117 
00118     for (i = 0; i < n; i++) {
00119         sum = 0;
00120         for (k = 0; k < n/2; k++) {
00121             a = (2 * i + 1 + (n / 2)) * (2 * k + 1);
00122             f = cos(M_PI * a / (double)(2 * n));
00123             sum += f * in[k];
00124         }
00125         out[i] = REF_SCALE(-sum, nbits - 2);
00126     }
00127 }
00128 
00129 /* NOTE: no normalisation by 1 / N is done */
00130 static void mdct_ref(FFTSample *output, FFTSample *input, int nbits)
00131 {
00132     int n = 1<<nbits;
00133     int k, i;
00134     double a, s;
00135 
00136     /* do it by hand */
00137     for (k = 0; k < n/2; k++) {
00138         s = 0;
00139         for (i = 0; i < n; i++) {
00140             a = (2*M_PI*(2*i+1+n/2)*(2*k+1) / (4 * n));
00141             s += input[i] * cos(a);
00142         }
00143         output[k] = REF_SCALE(s, nbits - 1);
00144     }
00145 }
00146 
00147 #if CONFIG_FFT_FLOAT
00148 static void idct_ref(float *output, float *input, int nbits)
00149 {
00150     int n = 1<<nbits;
00151     int k, i;
00152     double a, s;
00153 
00154     /* do it by hand */
00155     for (i = 0; i < n; i++) {
00156         s = 0.5 * input[0];
00157         for (k = 1; k < n; k++) {
00158             a = M_PI*k*(i+0.5) / n;
00159             s += input[k] * cos(a);
00160         }
00161         output[i] = 2 * s / n;
00162     }
00163 }
00164 static void dct_ref(float *output, float *input, int nbits)
00165 {
00166     int n = 1<<nbits;
00167     int k, i;
00168     double a, s;
00169 
00170     /* do it by hand */
00171     for (k = 0; k < n; k++) {
00172         s = 0;
00173         for (i = 0; i < n; i++) {
00174             a = M_PI*k*(i+0.5) / n;
00175             s += input[i] * cos(a);
00176         }
00177         output[k] = s;
00178     }
00179 }
00180 #endif
00181 
00182 
00183 static FFTSample frandom(AVLFG *prng)
00184 {
00185     return (int16_t)av_lfg_get(prng) / 32768.0 * RANGE;
00186 }
00187 
00188 static int64_t gettime(void)
00189 {
00190     struct timeval tv;
00191     gettimeofday(&tv,NULL);
00192     return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
00193 }
00194 
00195 static int check_diff(FFTSample *tab1, FFTSample *tab2, int n, double scale)
00196 {
00197     int i;
00198     double max= 0;
00199     double error= 0;
00200     int err = 0;
00201 
00202     for (i = 0; i < n; i++) {
00203         double e = fabsf(tab1[i] - (tab2[i] / scale)) / RANGE;
00204         if (e >= 1e-3) {
00205             av_log(NULL, AV_LOG_ERROR, "ERROR %5d: "FMT" "FMT"\n",
00206                    i, tab1[i], tab2[i]);
00207             err = 1;
00208         }
00209         error+= e*e;
00210         if(e>max) max= e;
00211     }
00212     av_log(NULL, AV_LOG_INFO, "max:%f e:%g\n", max, sqrt(error)/n);
00213     return err;
00214 }
00215 
00216 
00217 static void help(void)
00218 {
00219     av_log(NULL, AV_LOG_INFO,"usage: fft-test [-h] [-s] [-i] [-n b]\n"
00220            "-h     print this help\n"
00221            "-s     speed test\n"
00222            "-m     (I)MDCT test\n"
00223            "-d     (I)DCT test\n"
00224            "-r     (I)RDFT test\n"
00225            "-i     inverse transform test\n"
00226            "-n b   set the transform size to 2^b\n"
00227            "-f x   set scale factor for output data of (I)MDCT to x\n"
00228            );
00229 }
00230 
00231 enum tf_transform {
00232     TRANSFORM_FFT,
00233     TRANSFORM_MDCT,
00234     TRANSFORM_RDFT,
00235     TRANSFORM_DCT,
00236 };
00237 
00238 int main(int argc, char **argv)
00239 {
00240     FFTComplex *tab, *tab1, *tab_ref;
00241     FFTSample *tab2;
00242     int it, i, c;
00243     int do_speed = 0;
00244     int err = 1;
00245     enum tf_transform transform = TRANSFORM_FFT;
00246     int do_inverse = 0;
00247     FFTContext s1, *s = &s1;
00248     FFTContext m1, *m = &m1;
00249 #if CONFIG_FFT_FLOAT
00250     RDFTContext r1, *r = &r1;
00251     DCTContext d1, *d = &d1;
00252     int fft_size_2;
00253 #endif
00254     int fft_nbits, fft_size;
00255     double scale = 1.0;
00256     AVLFG prng;
00257     av_lfg_init(&prng, 1);
00258 
00259     fft_nbits = 9;
00260     for(;;) {
00261         c = getopt(argc, argv, "hsimrdn:f:");
00262         if (c == -1)
00263             break;
00264         switch(c) {
00265         case 'h':
00266             help();
00267             return 1;
00268         case 's':
00269             do_speed = 1;
00270             break;
00271         case 'i':
00272             do_inverse = 1;
00273             break;
00274         case 'm':
00275             transform = TRANSFORM_MDCT;
00276             break;
00277         case 'r':
00278             transform = TRANSFORM_RDFT;
00279             break;
00280         case 'd':
00281             transform = TRANSFORM_DCT;
00282             break;
00283         case 'n':
00284             fft_nbits = atoi(optarg);
00285             break;
00286         case 'f':
00287             scale = atof(optarg);
00288             break;
00289         }
00290     }
00291 
00292     fft_size = 1 << fft_nbits;
00293     tab = av_malloc(fft_size * sizeof(FFTComplex));
00294     tab1 = av_malloc(fft_size * sizeof(FFTComplex));
00295     tab_ref = av_malloc(fft_size * sizeof(FFTComplex));
00296     tab2 = av_malloc(fft_size * sizeof(FFTSample));
00297 
00298     switch (transform) {
00299     case TRANSFORM_MDCT:
00300         av_log(NULL, AV_LOG_INFO,"Scale factor is set to %f\n", scale);
00301         if (do_inverse)
00302             av_log(NULL, AV_LOG_INFO,"IMDCT");
00303         else
00304             av_log(NULL, AV_LOG_INFO,"MDCT");
00305         ff_mdct_init(m, fft_nbits, do_inverse, scale);
00306         break;
00307     case TRANSFORM_FFT:
00308         if (do_inverse)
00309             av_log(NULL, AV_LOG_INFO,"IFFT");
00310         else
00311             av_log(NULL, AV_LOG_INFO,"FFT");
00312         ff_fft_init(s, fft_nbits, do_inverse);
00313         fft_ref_init(fft_nbits, do_inverse);
00314         break;
00315 #if CONFIG_FFT_FLOAT
00316     case TRANSFORM_RDFT:
00317         if (do_inverse)
00318             av_log(NULL, AV_LOG_INFO,"IDFT_C2R");
00319         else
00320             av_log(NULL, AV_LOG_INFO,"DFT_R2C");
00321         ff_rdft_init(r, fft_nbits, do_inverse ? IDFT_C2R : DFT_R2C);
00322         fft_ref_init(fft_nbits, do_inverse);
00323         break;
00324     case TRANSFORM_DCT:
00325         if (do_inverse)
00326             av_log(NULL, AV_LOG_INFO,"DCT_III");
00327         else
00328             av_log(NULL, AV_LOG_INFO,"DCT_II");
00329         ff_dct_init(d, fft_nbits, do_inverse ? DCT_III : DCT_II);
00330         break;
00331 #endif
00332     default:
00333         av_log(NULL, AV_LOG_ERROR, "Requested transform not supported\n");
00334         return 1;
00335     }
00336     av_log(NULL, AV_LOG_INFO," %d test\n", fft_size);
00337 
00338     /* generate random data */
00339 
00340     for (i = 0; i < fft_size; i++) {
00341         tab1[i].re = frandom(&prng);
00342         tab1[i].im = frandom(&prng);
00343     }
00344 
00345     /* checking result */
00346     av_log(NULL, AV_LOG_INFO,"Checking...\n");
00347 
00348     switch (transform) {
00349     case TRANSFORM_MDCT:
00350         if (do_inverse) {
00351             imdct_ref((FFTSample *)tab_ref, (FFTSample *)tab1, fft_nbits);
00352             m->imdct_calc(m, tab2, (FFTSample *)tab1);
00353             err = check_diff((FFTSample *)tab_ref, tab2, fft_size, scale);
00354         } else {
00355             mdct_ref((FFTSample *)tab_ref, (FFTSample *)tab1, fft_nbits);
00356 
00357             m->mdct_calc(m, tab2, (FFTSample *)tab1);
00358 
00359             err = check_diff((FFTSample *)tab_ref, tab2, fft_size / 2, scale);
00360         }
00361         break;
00362     case TRANSFORM_FFT:
00363         memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
00364         s->fft_permute(s, tab);
00365         s->fft_calc(s, tab);
00366 
00367         fft_ref(tab_ref, tab1, fft_nbits);
00368         err = check_diff((FFTSample *)tab_ref, (FFTSample *)tab, fft_size * 2, 1.0);
00369         break;
00370 #if CONFIG_FFT_FLOAT
00371     case TRANSFORM_RDFT:
00372         fft_size_2 = fft_size >> 1;
00373         if (do_inverse) {
00374             tab1[         0].im = 0;
00375             tab1[fft_size_2].im = 0;
00376             for (i = 1; i < fft_size_2; i++) {
00377                 tab1[fft_size_2+i].re =  tab1[fft_size_2-i].re;
00378                 tab1[fft_size_2+i].im = -tab1[fft_size_2-i].im;
00379             }
00380 
00381             memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
00382             tab2[1] = tab1[fft_size_2].re;
00383 
00384             r->rdft_calc(r, tab2);
00385             fft_ref(tab_ref, tab1, fft_nbits);
00386             for (i = 0; i < fft_size; i++) {
00387                 tab[i].re = tab2[i];
00388                 tab[i].im = 0;
00389             }
00390             err = check_diff((float *)tab_ref, (float *)tab, fft_size * 2, 0.5);
00391         } else {
00392             for (i = 0; i < fft_size; i++) {
00393                 tab2[i]    = tab1[i].re;
00394                 tab1[i].im = 0;
00395             }
00396             r->rdft_calc(r, tab2);
00397             fft_ref(tab_ref, tab1, fft_nbits);
00398             tab_ref[0].im = tab_ref[fft_size_2].re;
00399             err = check_diff((float *)tab_ref, (float *)tab2, fft_size, 1.0);
00400         }
00401         break;
00402     case TRANSFORM_DCT:
00403         memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
00404         d->dct_calc(d, tab);
00405         if (do_inverse) {
00406             idct_ref(tab_ref, tab1, fft_nbits);
00407         } else {
00408             dct_ref(tab_ref, tab1, fft_nbits);
00409         }
00410         err = check_diff((float *)tab_ref, (float *)tab, fft_size, 1.0);
00411         break;
00412 #endif
00413     }
00414 
00415     /* do a speed test */
00416 
00417     if (do_speed) {
00418         int64_t time_start, duration;
00419         int nb_its;
00420 
00421         av_log(NULL, AV_LOG_INFO,"Speed test...\n");
00422         /* we measure during about 1 seconds */
00423         nb_its = 1;
00424         for(;;) {
00425             time_start = gettime();
00426             for (it = 0; it < nb_its; it++) {
00427                 switch (transform) {
00428                 case TRANSFORM_MDCT:
00429                     if (do_inverse) {
00430                         m->imdct_calc(m, (FFTSample *)tab, (FFTSample *)tab1);
00431                     } else {
00432                         m->mdct_calc(m, (FFTSample *)tab, (FFTSample *)tab1);
00433                     }
00434                     break;
00435                 case TRANSFORM_FFT:
00436                     memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
00437                     s->fft_calc(s, tab);
00438                     break;
00439 #if CONFIG_FFT_FLOAT
00440                 case TRANSFORM_RDFT:
00441                     memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
00442                     r->rdft_calc(r, tab2);
00443                     break;
00444                 case TRANSFORM_DCT:
00445                     memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
00446                     d->dct_calc(d, tab2);
00447                     break;
00448 #endif
00449                 }
00450             }
00451             duration = gettime() - time_start;
00452             if (duration >= 1000000)
00453                 break;
00454             nb_its *= 2;
00455         }
00456         av_log(NULL, AV_LOG_INFO,"time: %0.1f us/transform [total time=%0.2f s its=%d]\n",
00457                (double)duration / nb_its,
00458                (double)duration / 1000000.0,
00459                nb_its);
00460     }
00461 
00462     switch (transform) {
00463     case TRANSFORM_MDCT:
00464         ff_mdct_end(m);
00465         break;
00466     case TRANSFORM_FFT:
00467         ff_fft_end(s);
00468         break;
00469 #if CONFIG_FFT_FLOAT
00470     case TRANSFORM_RDFT:
00471         ff_rdft_end(r);
00472         break;
00473     case TRANSFORM_DCT:
00474         ff_dct_end(d);
00475         break;
00476 #endif
00477     }
00478 
00479     av_free(tab);
00480     av_free(tab1);
00481     av_free(tab2);
00482     av_free(tab_ref);
00483     av_free(exptab);
00484 
00485     return err;
00486 }