00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00026 #include "libavutil/cpu.h"
00027 #include "libavutil/mathematics.h"
00028 #include "libavutil/lfg.h"
00029 #include "libavutil/log.h"
00030 #include "fft.h"
00031 #if CONFIG_FFT_FLOAT
00032 #include "dct.h"
00033 #include "rdft.h"
00034 #endif
00035 #include <math.h>
00036 #include <unistd.h>
00037 #include <sys/time.h>
00038 #include <stdlib.h>
00039 #include <string.h>
00040
00041
00042
00043 #define MUL16(a,b) ((a) * (b))
00044
00045 #define CMAC(pre, pim, are, aim, bre, bim) \
00046 {\
00047 pre += (MUL16(are, bre) - MUL16(aim, bim));\
00048 pim += (MUL16(are, bim) + MUL16(bre, aim));\
00049 }
00050
00051 #if CONFIG_FFT_FLOAT
00052 # define RANGE 1.0
00053 # define REF_SCALE(x, bits) (x)
00054 # define FMT "%10.6f"
00055 #else
00056 # define RANGE 16384
00057 # define REF_SCALE(x, bits) ((x) / (1<<(bits)))
00058 # define FMT "%6d"
00059 #endif
00060
00061 struct {
00062 float re, im;
00063 } *exptab;
00064
00065 static void fft_ref_init(int nbits, int inverse)
00066 {
00067 int n, i;
00068 double c1, s1, alpha;
00069
00070 n = 1 << nbits;
00071 exptab = av_malloc((n / 2) * sizeof(*exptab));
00072
00073 for (i = 0; i < (n/2); i++) {
00074 alpha = 2 * M_PI * (float)i / (float)n;
00075 c1 = cos(alpha);
00076 s1 = sin(alpha);
00077 if (!inverse)
00078 s1 = -s1;
00079 exptab[i].re = c1;
00080 exptab[i].im = s1;
00081 }
00082 }
00083
00084 static void fft_ref(FFTComplex *tabr, FFTComplex *tab, int nbits)
00085 {
00086 int n, i, j, k, n2;
00087 double tmp_re, tmp_im, s, c;
00088 FFTComplex *q;
00089
00090 n = 1 << nbits;
00091 n2 = n >> 1;
00092 for (i = 0; i < n; i++) {
00093 tmp_re = 0;
00094 tmp_im = 0;
00095 q = tab;
00096 for (j = 0; j < n; j++) {
00097 k = (i * j) & (n - 1);
00098 if (k >= n2) {
00099 c = -exptab[k - n2].re;
00100 s = -exptab[k - n2].im;
00101 } else {
00102 c = exptab[k].re;
00103 s = exptab[k].im;
00104 }
00105 CMAC(tmp_re, tmp_im, c, s, q->re, q->im);
00106 q++;
00107 }
00108 tabr[i].re = REF_SCALE(tmp_re, nbits);
00109 tabr[i].im = REF_SCALE(tmp_im, nbits);
00110 }
00111 }
00112
00113 static void imdct_ref(FFTSample *out, FFTSample *in, int nbits)
00114 {
00115 int n = 1<<nbits;
00116 int k, i, a;
00117 double sum, f;
00118
00119 for (i = 0; i < n; i++) {
00120 sum = 0;
00121 for (k = 0; k < n/2; k++) {
00122 a = (2 * i + 1 + (n / 2)) * (2 * k + 1);
00123 f = cos(M_PI * a / (double)(2 * n));
00124 sum += f * in[k];
00125 }
00126 out[i] = REF_SCALE(-sum, nbits - 2);
00127 }
00128 }
00129
00130
00131 static void mdct_ref(FFTSample *output, FFTSample *input, int nbits)
00132 {
00133 int n = 1<<nbits;
00134 int k, i;
00135 double a, s;
00136
00137
00138 for (k = 0; k < n/2; k++) {
00139 s = 0;
00140 for (i = 0; i < n; i++) {
00141 a = (2*M_PI*(2*i+1+n/2)*(2*k+1) / (4 * n));
00142 s += input[i] * cos(a);
00143 }
00144 output[k] = REF_SCALE(s, nbits - 1);
00145 }
00146 }
00147
00148 #if CONFIG_FFT_FLOAT
00149 static void idct_ref(FFTSample *output, FFTSample *input, int nbits)
00150 {
00151 int n = 1<<nbits;
00152 int k, i;
00153 double a, s;
00154
00155
00156 for (i = 0; i < n; i++) {
00157 s = 0.5 * input[0];
00158 for (k = 1; k < n; k++) {
00159 a = M_PI*k*(i+0.5) / n;
00160 s += input[k] * cos(a);
00161 }
00162 output[i] = 2 * s / n;
00163 }
00164 }
00165 static void dct_ref(FFTSample *output, FFTSample *input, int nbits)
00166 {
00167 int n = 1<<nbits;
00168 int k, i;
00169 double a, s;
00170
00171
00172 for (k = 0; k < n; k++) {
00173 s = 0;
00174 for (i = 0; i < n; i++) {
00175 a = M_PI*k*(i+0.5) / n;
00176 s += input[i] * cos(a);
00177 }
00178 output[k] = s;
00179 }
00180 }
00181 #endif
00182
00183
00184 static FFTSample frandom(AVLFG *prng)
00185 {
00186 return (int16_t)av_lfg_get(prng) / 32768.0 * RANGE;
00187 }
00188
00189 static int64_t gettime(void)
00190 {
00191 struct timeval tv;
00192 gettimeofday(&tv,NULL);
00193 return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
00194 }
00195
00196 static int check_diff(FFTSample *tab1, FFTSample *tab2, int n, double scale)
00197 {
00198 int i;
00199 double max= 0;
00200 double error= 0;
00201 int err = 0;
00202
00203 for (i = 0; i < n; i++) {
00204 double e = fabsf(tab1[i] - (tab2[i] / scale)) / RANGE;
00205 if (e >= 1e-3) {
00206 av_log(NULL, AV_LOG_ERROR, "ERROR %5d: "FMT" "FMT"\n",
00207 i, tab1[i], tab2[i]);
00208 err = 1;
00209 }
00210 error+= e*e;
00211 if(e>max) max= e;
00212 }
00213 av_log(NULL, AV_LOG_INFO, "max:%f e:%g\n", max, sqrt(error)/n);
00214 return err;
00215 }
00216
00217
00218 static void help(void)
00219 {
00220 av_log(NULL, AV_LOG_INFO,"usage: fft-test [-h] [-s] [-i] [-n b]\n"
00221 "-h print this help\n"
00222 "-s speed test\n"
00223 "-m (I)MDCT test\n"
00224 "-d (I)DCT test\n"
00225 "-r (I)RDFT test\n"
00226 "-i inverse transform test\n"
00227 "-n b set the transform size to 2^b\n"
00228 "-f x set scale factor for output data of (I)MDCT to x\n"
00229 );
00230 }
00231
00232 enum tf_transform {
00233 TRANSFORM_FFT,
00234 TRANSFORM_MDCT,
00235 TRANSFORM_RDFT,
00236 TRANSFORM_DCT,
00237 };
00238
00239 int main(int argc, char **argv)
00240 {
00241 FFTComplex *tab, *tab1, *tab_ref;
00242 FFTSample *tab2;
00243 int it, i, c;
00244 int cpuflags;
00245 int do_speed = 0;
00246 int err = 1;
00247 enum tf_transform transform = TRANSFORM_FFT;
00248 int do_inverse = 0;
00249 FFTContext s1, *s = &s1;
00250 FFTContext m1, *m = &m1;
00251 #if CONFIG_FFT_FLOAT
00252 RDFTContext r1, *r = &r1;
00253 DCTContext d1, *d = &d1;
00254 int fft_size_2;
00255 #endif
00256 int fft_nbits, fft_size;
00257 double scale = 1.0;
00258 AVLFG prng;
00259 av_lfg_init(&prng, 1);
00260
00261 fft_nbits = 9;
00262 for(;;) {
00263 c = getopt(argc, argv, "hsimrdn:f:c:");
00264 if (c == -1)
00265 break;
00266 switch(c) {
00267 case 'h':
00268 help();
00269 return 1;
00270 case 's':
00271 do_speed = 1;
00272 break;
00273 case 'i':
00274 do_inverse = 1;
00275 break;
00276 case 'm':
00277 transform = TRANSFORM_MDCT;
00278 break;
00279 case 'r':
00280 transform = TRANSFORM_RDFT;
00281 break;
00282 case 'd':
00283 transform = TRANSFORM_DCT;
00284 break;
00285 case 'n':
00286 fft_nbits = atoi(optarg);
00287 break;
00288 case 'f':
00289 scale = atof(optarg);
00290 break;
00291 case 'c':
00292 cpuflags = av_parse_cpu_flags(optarg);
00293 if (cpuflags < 0)
00294 return 1;
00295 av_set_cpu_flags_mask(cpuflags);
00296 break;
00297 }
00298 }
00299
00300 fft_size = 1 << fft_nbits;
00301 tab = av_malloc(fft_size * sizeof(FFTComplex));
00302 tab1 = av_malloc(fft_size * sizeof(FFTComplex));
00303 tab_ref = av_malloc(fft_size * sizeof(FFTComplex));
00304 tab2 = av_malloc(fft_size * sizeof(FFTSample));
00305
00306 switch (transform) {
00307 case TRANSFORM_MDCT:
00308 av_log(NULL, AV_LOG_INFO,"Scale factor is set to %f\n", scale);
00309 if (do_inverse)
00310 av_log(NULL, AV_LOG_INFO,"IMDCT");
00311 else
00312 av_log(NULL, AV_LOG_INFO,"MDCT");
00313 ff_mdct_init(m, fft_nbits, do_inverse, scale);
00314 break;
00315 case TRANSFORM_FFT:
00316 if (do_inverse)
00317 av_log(NULL, AV_LOG_INFO,"IFFT");
00318 else
00319 av_log(NULL, AV_LOG_INFO,"FFT");
00320 ff_fft_init(s, fft_nbits, do_inverse);
00321 fft_ref_init(fft_nbits, do_inverse);
00322 break;
00323 #if CONFIG_FFT_FLOAT
00324 case TRANSFORM_RDFT:
00325 if (do_inverse)
00326 av_log(NULL, AV_LOG_INFO,"IDFT_C2R");
00327 else
00328 av_log(NULL, AV_LOG_INFO,"DFT_R2C");
00329 ff_rdft_init(r, fft_nbits, do_inverse ? IDFT_C2R : DFT_R2C);
00330 fft_ref_init(fft_nbits, do_inverse);
00331 break;
00332 case TRANSFORM_DCT:
00333 if (do_inverse)
00334 av_log(NULL, AV_LOG_INFO,"DCT_III");
00335 else
00336 av_log(NULL, AV_LOG_INFO,"DCT_II");
00337 ff_dct_init(d, fft_nbits, do_inverse ? DCT_III : DCT_II);
00338 break;
00339 #endif
00340 default:
00341 av_log(NULL, AV_LOG_ERROR, "Requested transform not supported\n");
00342 return 1;
00343 }
00344 av_log(NULL, AV_LOG_INFO," %d test\n", fft_size);
00345
00346
00347
00348 for (i = 0; i < fft_size; i++) {
00349 tab1[i].re = frandom(&prng);
00350 tab1[i].im = frandom(&prng);
00351 }
00352
00353
00354 av_log(NULL, AV_LOG_INFO,"Checking...\n");
00355
00356 switch (transform) {
00357 case TRANSFORM_MDCT:
00358 if (do_inverse) {
00359 imdct_ref((FFTSample *)tab_ref, (FFTSample *)tab1, fft_nbits);
00360 m->imdct_calc(m, tab2, (FFTSample *)tab1);
00361 err = check_diff((FFTSample *)tab_ref, tab2, fft_size, scale);
00362 } else {
00363 mdct_ref((FFTSample *)tab_ref, (FFTSample *)tab1, fft_nbits);
00364
00365 m->mdct_calc(m, tab2, (FFTSample *)tab1);
00366
00367 err = check_diff((FFTSample *)tab_ref, tab2, fft_size / 2, scale);
00368 }
00369 break;
00370 case TRANSFORM_FFT:
00371 memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
00372 s->fft_permute(s, tab);
00373 s->fft_calc(s, tab);
00374
00375 fft_ref(tab_ref, tab1, fft_nbits);
00376 err = check_diff((FFTSample *)tab_ref, (FFTSample *)tab, fft_size * 2, 1.0);
00377 break;
00378 #if CONFIG_FFT_FLOAT
00379 case TRANSFORM_RDFT:
00380 fft_size_2 = fft_size >> 1;
00381 if (do_inverse) {
00382 tab1[ 0].im = 0;
00383 tab1[fft_size_2].im = 0;
00384 for (i = 1; i < fft_size_2; i++) {
00385 tab1[fft_size_2+i].re = tab1[fft_size_2-i].re;
00386 tab1[fft_size_2+i].im = -tab1[fft_size_2-i].im;
00387 }
00388
00389 memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
00390 tab2[1] = tab1[fft_size_2].re;
00391
00392 r->rdft_calc(r, tab2);
00393 fft_ref(tab_ref, tab1, fft_nbits);
00394 for (i = 0; i < fft_size; i++) {
00395 tab[i].re = tab2[i];
00396 tab[i].im = 0;
00397 }
00398 err = check_diff((float *)tab_ref, (float *)tab, fft_size * 2, 0.5);
00399 } else {
00400 for (i = 0; i < fft_size; i++) {
00401 tab2[i] = tab1[i].re;
00402 tab1[i].im = 0;
00403 }
00404 r->rdft_calc(r, tab2);
00405 fft_ref(tab_ref, tab1, fft_nbits);
00406 tab_ref[0].im = tab_ref[fft_size_2].re;
00407 err = check_diff((float *)tab_ref, (float *)tab2, fft_size, 1.0);
00408 }
00409 break;
00410 case TRANSFORM_DCT:
00411 memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
00412 d->dct_calc(d, (FFTSample *)tab);
00413 if (do_inverse) {
00414 idct_ref((FFTSample*)tab_ref, (FFTSample *)tab1, fft_nbits);
00415 } else {
00416 dct_ref((FFTSample*)tab_ref, (FFTSample *)tab1, fft_nbits);
00417 }
00418 err = check_diff((float *)tab_ref, (float *)tab, fft_size, 1.0);
00419 break;
00420 #endif
00421 }
00422
00423
00424
00425 if (do_speed) {
00426 int64_t time_start, duration;
00427 int nb_its;
00428
00429 av_log(NULL, AV_LOG_INFO,"Speed test...\n");
00430
00431 nb_its = 1;
00432 for(;;) {
00433 time_start = gettime();
00434 for (it = 0; it < nb_its; it++) {
00435 switch (transform) {
00436 case TRANSFORM_MDCT:
00437 if (do_inverse) {
00438 m->imdct_calc(m, (FFTSample *)tab, (FFTSample *)tab1);
00439 } else {
00440 m->mdct_calc(m, (FFTSample *)tab, (FFTSample *)tab1);
00441 }
00442 break;
00443 case TRANSFORM_FFT:
00444 memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
00445 s->fft_calc(s, tab);
00446 break;
00447 #if CONFIG_FFT_FLOAT
00448 case TRANSFORM_RDFT:
00449 memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
00450 r->rdft_calc(r, tab2);
00451 break;
00452 case TRANSFORM_DCT:
00453 memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
00454 d->dct_calc(d, tab2);
00455 break;
00456 #endif
00457 }
00458 }
00459 duration = gettime() - time_start;
00460 if (duration >= 1000000)
00461 break;
00462 nb_its *= 2;
00463 }
00464 av_log(NULL, AV_LOG_INFO,"time: %0.1f us/transform [total time=%0.2f s its=%d]\n",
00465 (double)duration / nb_its,
00466 (double)duration / 1000000.0,
00467 nb_its);
00468 }
00469
00470 switch (transform) {
00471 case TRANSFORM_MDCT:
00472 ff_mdct_end(m);
00473 break;
00474 case TRANSFORM_FFT:
00475 ff_fft_end(s);
00476 break;
00477 #if CONFIG_FFT_FLOAT
00478 case TRANSFORM_RDFT:
00479 ff_rdft_end(r);
00480 break;
00481 case TRANSFORM_DCT:
00482 ff_dct_end(d);
00483 break;
00484 #endif
00485 }
00486
00487 av_free(tab);
00488 av_free(tab1);
00489 av_free(tab2);
00490 av_free(tab_ref);
00491 av_free(exptab);
00492
00493 return err;
00494 }