libavcodec/fft-test.c

   1 /*
   2  * (c) 2002 Fabrice Bellard
   3  *
   4  * This file is part of Libav.
   5  *
   6  * Libav is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * Libav is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with Libav; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 /**
  22  * @file
  23  * FFT and MDCT tests.
  24  */
  25
  26 #include "libavutil/cpu.h"
  27 #include "libavutil/mathematics.h"
  28 #include "libavutil/lfg.h"
  29 #include "libavutil/log.h"
  30 #include "libavutil/time.h"
  31 #include "fft.h"
  32 #if CONFIG_FFT_FLOAT
  33 #include "dct.h"
  34 #include "rdft.h"
  35 #endif
  36 #include <math.h>
  37 #if HAVE_UNISTD_H
  38 #include <unistd.h>
  39 #endif
  40 #include <stdlib.h>
  41 #include <string.h>
  42
  43 /* reference fft */
  44
  45 #define MUL16(a,b) ((a) * (b))
  46
  47 #define CMAC(pre, pim, are, aim, bre, bim) \
  48 {\
  49    pre += (MUL16(are, bre) - MUL16(aim, bim));\
  50    pim += (MUL16(are, bim) + MUL16(bre, aim));\
  51 }
  52
  53 #if CONFIG_FFT_FLOAT
  54 #   define RANGE 1.0
  55 #   define REF_SCALE(x, bits)  (x)
  56 #   define FMT "%10.6f"
  57 #else
  58 #   define RANGE 16384
  59 #   define REF_SCALE(x, bits) ((x) / (1<<(bits)))
  60 #   define FMT "%6d"
  61 #endif
  62
  63 struct {
  64     float re, im;
  65 } *exptab;
  66
  67 static void fft_ref_init(int nbits, int inverse)
  68 {
  69     int n, i;
  70     double c1, s1, alpha;
  71
  72     n = 1 << nbits;
  73     exptab = av_malloc((n / 2) * sizeof(*exptab));
  74
  75     for (i = 0; i < (n/2); i++) {
  76         alpha = 2 * M_PI * (float)i / (float)n;
  77         c1 = cos(alpha);
  78         s1 = sin(alpha);
  79         if (!inverse)
  80             s1 = -s1;
  81         exptab[i].re = c1;
  82         exptab[i].im = s1;
  83     }
  84 }
  85
  86 static void fft_ref(FFTComplex *tabr, FFTComplex *tab, int nbits)
  87 {
  88     int n, i, j, k, n2;
  89     double tmp_re, tmp_im, s, c;
  90     FFTComplex *q;
  91
  92     n = 1 << nbits;
  93     n2 = n >> 1;
  94     for (i = 0; i < n; i++) {
  95         tmp_re = 0;
  96         tmp_im = 0;
  97         q = tab;
  98         for (j = 0; j < n; j++) {
  99             k = (i * j) & (n - 1);
 100             if (k >= n2) {
 101                 c = -exptab[k - n2].re;
 102                 s = -exptab[k - n2].im;
 103             } else {
 104                 c = exptab[k].re;
 105                 s = exptab[k].im;
 106             }
 107             CMAC(tmp_re, tmp_im, c, s, q->re, q->im);
 108             q++;
 109         }
 110         tabr[i].re = REF_SCALE(tmp_re, nbits);
 111         tabr[i].im = REF_SCALE(tmp_im, nbits);
 112     }
 113 }
 114
 115 static void imdct_ref(FFTSample *out, FFTSample *in, int nbits)
 116 {
 117     int n = 1<<nbits;
 118     int k, i, a;
 119     double sum, f;
 120
 121     for (i = 0; i < n; i++) {
 122         sum = 0;
 123         for (k = 0; k < n/2; k++) {
 124             a = (2 * i + 1 + (n / 2)) * (2 * k + 1);
 125             f = cos(M_PI * a / (double)(2 * n));
 126             sum += f * in[k];
 127         }
 128         out[i] = REF_SCALE(-sum, nbits - 2);
 129     }
 130 }
 131
 132 /* NOTE: no normalisation by 1 / N is done */
 133 static void mdct_ref(FFTSample *output, FFTSample *input, int nbits)
 134 {
 135     int n = 1<<nbits;
 136     int k, i;
 137     double a, s;
 138
 139     /* do it by hand */
 140     for (k = 0; k < n/2; k++) {
 141         s = 0;
 142         for (i = 0; i < n; i++) {
 143             a = (2*M_PI*(2*i+1+n/2)*(2*k+1) / (4 * n));
 144             s += input[i] * cos(a);
 145         }
 146         output[k] = REF_SCALE(s, nbits - 1);
 147     }
 148 }
 149
 150 #if CONFIG_FFT_FLOAT
 151 static void idct_ref(float *output, float *input, int nbits)
 152 {
 153     int n = 1<<nbits;
 154     int k, i;
 155     double a, s;
 156
 157     /* do it by hand */
 158     for (i = 0; i < n; i++) {
 159         s = 0.5 * input[0];
 160         for (k = 1; k < n; k++) {
 161             a = M_PI*k*(i+0.5) / n;
 162             s += input[k] * cos(a);
 163         }
 164         output[i] = 2 * s / n;
 165     }
 166 }
 167 static void dct_ref(float *output, float *input, int nbits)
 168 {
 169     int n = 1<<nbits;
 170     int k, i;
 171     double a, s;
 172
 173     /* do it by hand */
 174     for (k = 0; k < n; k++) {
 175         s = 0;
 176         for (i = 0; i < n; i++) {
 177             a = M_PI*k*(i+0.5) / n;
 178             s += input[i] * cos(a);
 179         }
 180         output[k] = s;
 181     }
 182 }
 183 #endif
 184
 185
 186 static FFTSample frandom(AVLFG *prng)
 187 {
 188     return (int16_t)av_lfg_get(prng) / 32768.0 * RANGE;
 189 }
 190
 191 static int check_diff(FFTSample *tab1, FFTSample *tab2, int n, double scale)
 192 {
 193     int i;
 194     double max= 0;
 195     double error= 0;
 196     int err = 0;
 197
 198     for (i = 0; i < n; i++) {
 199         double e = fabsf(tab1[i] - (tab2[i] / scale)) / RANGE;
 200         if (e >= 1e-3) {
 201             av_log(NULL, AV_LOG_ERROR, "ERROR %5d: "FMT" "FMT"\n",
 202                    i, tab1[i], tab2[i]);
 203             err = 1;
 204         }
 205         error+= e*e;
 206         if(e>max) max= e;
 207     }
 208     av_log(NULL, AV_LOG_INFO, "max:%f e:%g\n", max, sqrt(error)/n);
 209     return err;
 210 }
 211
 212
 213 static void help(void)
 214 {
 215     av_log(NULL, AV_LOG_INFO,"usage: fft-test [-h] [-s] [-i] [-n b]\n"
 216            "-h     print this help\n"
 217            "-s     speed test\n"
 218            "-m     (I)MDCT test\n"
 219            "-d     (I)DCT test\n"
 220            "-r     (I)RDFT test\n"
 221            "-i     inverse transform test\n"
 222            "-n b   set the transform size to 2^b\n"
 223            "-f x   set scale factor for output data of (I)MDCT to x\n"
 224            );
 225 }
 226
 227 enum tf_transform {
 228     TRANSFORM_FFT,
 229     TRANSFORM_MDCT,
 230     TRANSFORM_RDFT,
 231     TRANSFORM_DCT,
 232 };
 233
 234 #if !HAVE_GETOPT
 235 #include "compat/getopt.c"
 236 #endif
 237
 238 int main(int argc, char **argv)
 239 {
 240     FFTComplex *tab, *tab1, *tab_ref;
 241     FFTSample *tab2;
 242     int it, i, c;
 243     int cpuflags;
 244     int do_speed = 0;
 245     int err = 1;
 246     enum tf_transform transform = TRANSFORM_FFT;
 247     int do_inverse = 0;
 248     FFTContext s1, *s = &s1;
 249     FFTContext m1, *m = &m1;
 250 #if CONFIG_FFT_FLOAT
 251     RDFTContext r1, *r = &r1;
 252     DCTContext d1, *d = &d1;
 253     int fft_size_2;
 254 #endif
 255     int fft_nbits, fft_size;
 256     double scale = 1.0;
 257     AVLFG prng;
 258     av_lfg_init(&prng, 1);
 259
 260     fft_nbits = 9;
 261     for(;;) {
 262         c = getopt(argc, argv, "hsimrdn:f:c:");
 263         if (c == -1)
 264             break;
 265         switch(c) {
 266         case 'h':
 267             help();
 268             return 1;
 269         case 's':
 270             do_speed = 1;
 271             break;
 272         case 'i':
 273             do_inverse = 1;
 274             break;
 275         case 'm':
 276             transform = TRANSFORM_MDCT;
 277             break;
 278         case 'r':
 279             transform = TRANSFORM_RDFT;
 280             break;
 281         case 'd':
 282             transform = TRANSFORM_DCT;
 283             break;
 284         case 'n':
 285             fft_nbits = atoi(optarg);
 286             break;
 287         case 'f':
 288             scale = atof(optarg);
 289             break;
 290         case 'c':
 291             cpuflags = av_parse_cpu_flags(optarg);
 292             if (cpuflags < 0)
 293                 return 1;
 294             av_set_cpu_flags_mask(cpuflags);
 295             break;
 296         }
 297     }
 298
 299     fft_size = 1 << fft_nbits;
 300     tab = av_malloc(fft_size * sizeof(FFTComplex));
 301     tab1 = av_malloc(fft_size * sizeof(FFTComplex));
 302     tab_ref = av_malloc(fft_size * sizeof(FFTComplex));
 303     tab2 = av_malloc(fft_size * sizeof(FFTSample));
 304
 305     switch (transform) {
 306     case TRANSFORM_MDCT:
 307         av_log(NULL, AV_LOG_INFO,"Scale factor is set to %f\n", scale);
 308         if (do_inverse)
 309             av_log(NULL, AV_LOG_INFO,"IMDCT");
 310         else
 311             av_log(NULL, AV_LOG_INFO,"MDCT");
 312         ff_mdct_init(m, fft_nbits, do_inverse, scale);
 313         break;
 314     case TRANSFORM_FFT:
 315         if (do_inverse)
 316             av_log(NULL, AV_LOG_INFO,"IFFT");
 317         else
 318             av_log(NULL, AV_LOG_INFO,"FFT");
 319         ff_fft_init(s, fft_nbits, do_inverse);
 320         fft_ref_init(fft_nbits, do_inverse);
 321         break;
 322 #if CONFIG_FFT_FLOAT
 323     case TRANSFORM_RDFT:
 324         if (do_inverse)
 325             av_log(NULL, AV_LOG_INFO,"IDFT_C2R");
 326         else
 327             av_log(NULL, AV_LOG_INFO,"DFT_R2C");
 328         ff_rdft_init(r, fft_nbits, do_inverse ? IDFT_C2R : DFT_R2C);
 329         fft_ref_init(fft_nbits, do_inverse);
 330         break;
 331     case TRANSFORM_DCT:
 332         if (do_inverse)
 333             av_log(NULL, AV_LOG_INFO,"DCT_III");
 334         else
 335             av_log(NULL, AV_LOG_INFO,"DCT_II");
 336         ff_dct_init(d, fft_nbits, do_inverse ? DCT_III : DCT_II);
 337         break;
 338 #endif
 339     default:
 340         av_log(NULL, AV_LOG_ERROR, "Requested transform not supported\n");
 341         return 1;
 342     }
 343     av_log(NULL, AV_LOG_INFO," %d test\n", fft_size);
 344
 345     /* generate random data */
 346
 347     for (i = 0; i < fft_size; i++) {
 348         tab1[i].re = frandom(&prng);
 349         tab1[i].im = frandom(&prng);
 350     }
 351
 352     /* checking result */
 353     av_log(NULL, AV_LOG_INFO,"Checking...\n");
 354
 355     switch (transform) {
 356     case TRANSFORM_MDCT:
 357         if (do_inverse) {
 358             imdct_ref((FFTSample *)tab_ref, (FFTSample *)tab1, fft_nbits);
 359             m->imdct_calc(m, tab2, (FFTSample *)tab1);
 360             err = check_diff((FFTSample *)tab_ref, tab2, fft_size, scale);
 361         } else {
 362             mdct_ref((FFTSample *)tab_ref, (FFTSample *)tab1, fft_nbits);
 363
 364             m->mdct_calc(m, tab2, (FFTSample *)tab1);
 365
 366             err = check_diff((FFTSample *)tab_ref, tab2, fft_size / 2, scale);
 367         }
 368         break;
 369     case TRANSFORM_FFT:
 370         memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
 371         s->fft_permute(s, tab);
 372         s->fft_calc(s, tab);
 373
 374         fft_ref(tab_ref, tab1, fft_nbits);
 375         err = check_diff((FFTSample *)tab_ref, (FFTSample *)tab, fft_size * 2, 1.0);
 376         break;
 377 #if CONFIG_FFT_FLOAT
 378     case TRANSFORM_RDFT:
 379         fft_size_2 = fft_size >> 1;
 380         if (do_inverse) {
 381             tab1[         0].im = 0;
 382             tab1[fft_size_2].im = 0;
 383             for (i = 1; i < fft_size_2; i++) {
 384                 tab1[fft_size_2+i].re =  tab1[fft_size_2-i].re;
 385                 tab1[fft_size_2+i].im = -tab1[fft_size_2-i].im;
 386             }
 387
 388             memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
 389             tab2[1] = tab1[fft_size_2].re;
 390
 391             r->rdft_calc(r, tab2);
 392             fft_ref(tab_ref, tab1, fft_nbits);
 393             for (i = 0; i < fft_size; i++) {
 394                 tab[i].re = tab2[i];
 395                 tab[i].im = 0;
 396             }
 397             err = check_diff((float *)tab_ref, (float *)tab, fft_size * 2, 0.5);
 398         } else {
 399             for (i = 0; i < fft_size; i++) {
 400                 tab2[i]    = tab1[i].re;
 401                 tab1[i].im = 0;
 402             }
 403             r->rdft_calc(r, tab2);
 404             fft_ref(tab_ref, tab1, fft_nbits);
 405             tab_ref[0].im = tab_ref[fft_size_2].re;
 406             err = check_diff((float *)tab_ref, (float *)tab2, fft_size, 1.0);
 407         }
 408         break;
 409     case TRANSFORM_DCT:
 410         memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
 411         d->dct_calc(d, tab);
 412         if (do_inverse) {
 413             idct_ref(tab_ref, tab1, fft_nbits);
 414         } else {
 415             dct_ref(tab_ref, tab1, fft_nbits);
 416         }
 417         err = check_diff((float *)tab_ref, (float *)tab, fft_size, 1.0);
 418         break;
 419 #endif
 420     }
 421
 422     /* do a speed test */
 423
 424     if (do_speed) {
 425         int64_t time_start, duration;
 426         int nb_its;
 427
 428         av_log(NULL, AV_LOG_INFO,"Speed test...\n");
 429         /* we measure during about 1 seconds */
 430         nb_its = 1;
 431         for(;;) {
 432             time_start = av_gettime();
 433             for (it = 0; it < nb_its; it++) {
 434                 switch (transform) {
 435                 case TRANSFORM_MDCT:
 436                     if (do_inverse) {
 437                         m->imdct_calc(m, (FFTSample *)tab, (FFTSample *)tab1);
 438                     } else {
 439                         m->mdct_calc(m, (FFTSample *)tab, (FFTSample *)tab1);
 440                     }
 441                     break;
 442                 case TRANSFORM_FFT:
 443                     memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
 444                     s->fft_calc(s, tab);
 445                     break;
 446 #if CONFIG_FFT_FLOAT
 447                 case TRANSFORM_RDFT:
 448                     memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
 449                     r->rdft_calc(r, tab2);
 450                     break;
 451                 case TRANSFORM_DCT:
 452                     memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
 453                     d->dct_calc(d, tab2);
 454                     break;
 455 #endif
 456                 }
 457             }
 458             duration = av_gettime() - time_start;
 459             if (duration >= 1000000)
 460                 break;
 461             nb_its *= 2;
 462         }
 463         av_log(NULL, AV_LOG_INFO,"time: %0.1f us/transform [total time=%0.2f s its=%d]\n",
 464                (double)duration / nb_its,
 465                (double)duration / 1000000.0,
 466                nb_its);
 467     }
 468
 469     switch (transform) {
 470     case TRANSFORM_MDCT:
 471         ff_mdct_end(m);
 472         break;
 473     case TRANSFORM_FFT:
 474         ff_fft_end(s);
 475         break;
 476 #if CONFIG_FFT_FLOAT
 477     case TRANSFORM_RDFT:
 478         ff_rdft_end(r);
 479         break;
 480     case TRANSFORM_DCT:
 481         ff_dct_end(d);
 482         break;
 483 #endif
 484     }
 485
 486     av_free(tab);
 487     av_free(tab1);
 488     av_free(tab2);
 489     av_free(tab_ref);
 490     av_free(exptab);
 491
 492     return err;
 493 }