1/*
2 * (c) 2002 Fabrice Bellard
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg 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 * FFmpeg 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 FFmpeg; 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/mathematics.h"
27#include "libavutil/lfg.h"
28#include "libavutil/log.h"
29#include "fft.h"
30#include <math.h>
31#include <unistd.h>
32#include <sys/time.h>
33#include <stdlib.h>
34#include <string.h>
35
36#undef exit
37
38/* reference fft */
39
40#define MUL16(a,b) ((a) * (b))
41
42#define CMAC(pre, pim, are, aim, bre, bim) \
43{\
44   pre += (MUL16(are, bre) - MUL16(aim, bim));\
45   pim += (MUL16(are, bim) + MUL16(bre, aim));\
46}
47
48FFTComplex *exptab;
49
50static void fft_ref_init(int nbits, int inverse)
51{
52    int n, i;
53    double c1, s1, alpha;
54
55    n = 1 << nbits;
56    exptab = av_malloc((n / 2) * sizeof(FFTComplex));
57
58    for (i = 0; i < (n/2); i++) {
59        alpha = 2 * M_PI * (float)i / (float)n;
60        c1 = cos(alpha);
61        s1 = sin(alpha);
62        if (!inverse)
63            s1 = -s1;
64        exptab[i].re = c1;
65        exptab[i].im = s1;
66    }
67}
68
69static void fft_ref(FFTComplex *tabr, FFTComplex *tab, int nbits)
70{
71    int n, i, j, k, n2;
72    double tmp_re, tmp_im, s, c;
73    FFTComplex *q;
74
75    n = 1 << nbits;
76    n2 = n >> 1;
77    for (i = 0; i < n; i++) {
78        tmp_re = 0;
79        tmp_im = 0;
80        q = tab;
81        for (j = 0; j < n; j++) {
82            k = (i * j) & (n - 1);
83            if (k >= n2) {
84                c = -exptab[k - n2].re;
85                s = -exptab[k - n2].im;
86            } else {
87                c = exptab[k].re;
88                s = exptab[k].im;
89            }
90            CMAC(tmp_re, tmp_im, c, s, q->re, q->im);
91            q++;
92        }
93        tabr[i].re = tmp_re;
94        tabr[i].im = tmp_im;
95    }
96}
97
98static void imdct_ref(float *out, float *in, int nbits)
99{
100    int n = 1<<nbits;
101    int k, i, a;
102    double sum, f;
103
104    for (i = 0; i < n; i++) {
105        sum = 0;
106        for (k = 0; k < n/2; k++) {
107            a = (2 * i + 1 + (n / 2)) * (2 * k + 1);
108            f = cos(M_PI * a / (double)(2 * n));
109            sum += f * in[k];
110        }
111        out[i] = -sum;
112    }
113}
114
115/* NOTE: no normalisation by 1 / N is done */
116static void mdct_ref(float *output, float *input, int nbits)
117{
118    int n = 1<<nbits;
119    int k, i;
120    double a, s;
121
122    /* do it by hand */
123    for (k = 0; k < n/2; k++) {
124        s = 0;
125        for (i = 0; i < n; i++) {
126            a = (2*M_PI*(2*i+1+n/2)*(2*k+1) / (4 * n));
127            s += input[i] * cos(a);
128        }
129        output[k] = s;
130    }
131}
132
133static void idct_ref(float *output, float *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 (i = 0; i < n; i++) {
141        s = 0.5 * input[0];
142        for (k = 1; k < n; k++) {
143            a = M_PI*k*(i+0.5) / n;
144            s += input[k] * cos(a);
145        }
146        output[i] = 2 * s / n;
147    }
148}
149static void dct_ref(float *output, float *input, int nbits)
150{
151    int n = 1<<nbits;
152    int k, i;
153    double a, s;
154
155    /* do it by hand */
156    for (k = 0; k < n; k++) {
157        s = 0;
158        for (i = 0; i < n; i++) {
159            a = M_PI*k*(i+0.5) / n;
160            s += input[i] * cos(a);
161        }
162        output[k] = s;
163    }
164}
165
166
167static float frandom(AVLFG *prng)
168{
169    return (int16_t)av_lfg_get(prng) / 32768.0;
170}
171
172static int64_t gettime(void)
173{
174    struct timeval tv;
175    gettimeofday(&tv,NULL);
176    return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
177}
178
179static void check_diff(float *tab1, float *tab2, int n, double scale)
180{
181    int i;
182    double max= 0;
183    double error= 0;
184
185    for (i = 0; i < n; i++) {
186        double e= fabsf(tab1[i] - (tab2[i] / scale));
187        if (e >= 1e-3) {
188            av_log(NULL, AV_LOG_ERROR, "ERROR %d: %f %f\n",
189                   i, tab1[i], tab2[i]);
190        }
191        error+= e*e;
192        if(e>max) max= e;
193    }
194    av_log(NULL, AV_LOG_INFO, "max:%f e:%g\n", max, sqrt(error)/n);
195}
196
197
198static void help(void)
199{
200    av_log(NULL, AV_LOG_INFO,"usage: fft-test [-h] [-s] [-i] [-n b]\n"
201           "-h     print this help\n"
202           "-s     speed test\n"
203           "-m     (I)MDCT test\n"
204           "-d     (I)DCT test\n"
205           "-r     (I)RDFT test\n"
206           "-i     inverse transform test\n"
207           "-n b   set the transform size to 2^b\n"
208           "-f x   set scale factor for output data of (I)MDCT to x\n"
209           );
210    exit(1);
211}
212
213enum tf_transform {
214    TRANSFORM_FFT,
215    TRANSFORM_MDCT,
216    TRANSFORM_RDFT,
217    TRANSFORM_DCT,
218};
219
220int main(int argc, char **argv)
221{
222    FFTComplex *tab, *tab1, *tab_ref;
223    FFTSample *tab2;
224    int it, i, c;
225    int do_speed = 0;
226    enum tf_transform transform = TRANSFORM_FFT;
227    int do_inverse = 0;
228    FFTContext s1, *s = &s1;
229    FFTContext m1, *m = &m1;
230    RDFTContext r1, *r = &r1;
231    DCTContext d1, *d = &d1;
232    int fft_nbits, fft_size, fft_size_2;
233    double scale = 1.0;
234    AVLFG prng;
235    av_lfg_init(&prng, 1);
236
237    fft_nbits = 9;
238    for(;;) {
239        c = getopt(argc, argv, "hsimrdn:f:");
240        if (c == -1)
241            break;
242        switch(c) {
243        case 'h':
244            help();
245            break;
246        case 's':
247            do_speed = 1;
248            break;
249        case 'i':
250            do_inverse = 1;
251            break;
252        case 'm':
253            transform = TRANSFORM_MDCT;
254            break;
255        case 'r':
256            transform = TRANSFORM_RDFT;
257            break;
258        case 'd':
259            transform = TRANSFORM_DCT;
260            break;
261        case 'n':
262            fft_nbits = atoi(optarg);
263            break;
264        case 'f':
265            scale = atof(optarg);
266            break;
267        }
268    }
269
270    fft_size = 1 << fft_nbits;
271    fft_size_2 = fft_size >> 1;
272    tab = av_malloc(fft_size * sizeof(FFTComplex));
273    tab1 = av_malloc(fft_size * sizeof(FFTComplex));
274    tab_ref = av_malloc(fft_size * sizeof(FFTComplex));
275    tab2 = av_malloc(fft_size * sizeof(FFTSample));
276
277    switch (transform) {
278    case TRANSFORM_MDCT:
279        av_log(NULL, AV_LOG_INFO,"Scale factor is set to %f\n", scale);
280        if (do_inverse)
281            av_log(NULL, AV_LOG_INFO,"IMDCT");
282        else
283            av_log(NULL, AV_LOG_INFO,"MDCT");
284        ff_mdct_init(m, fft_nbits, do_inverse, scale);
285        break;
286    case TRANSFORM_FFT:
287        if (do_inverse)
288            av_log(NULL, AV_LOG_INFO,"IFFT");
289        else
290            av_log(NULL, AV_LOG_INFO,"FFT");
291        ff_fft_init(s, fft_nbits, do_inverse);
292        fft_ref_init(fft_nbits, do_inverse);
293        break;
294    case TRANSFORM_RDFT:
295        if (do_inverse)
296            av_log(NULL, AV_LOG_INFO,"IDFT_C2R");
297        else
298            av_log(NULL, AV_LOG_INFO,"DFT_R2C");
299        ff_rdft_init(r, fft_nbits, do_inverse ? IDFT_C2R : DFT_R2C);
300        fft_ref_init(fft_nbits, do_inverse);
301        break;
302    case TRANSFORM_DCT:
303        if (do_inverse)
304            av_log(NULL, AV_LOG_INFO,"DCT_III");
305        else
306            av_log(NULL, AV_LOG_INFO,"DCT_II");
307        ff_dct_init(d, fft_nbits, do_inverse ? DCT_III : DCT_II);
308        break;
309    }
310    av_log(NULL, AV_LOG_INFO," %d test\n", fft_size);
311
312    /* generate random data */
313
314    for (i = 0; i < fft_size; i++) {
315        tab1[i].re = frandom(&prng);
316        tab1[i].im = frandom(&prng);
317    }
318
319    /* checking result */
320    av_log(NULL, AV_LOG_INFO,"Checking...\n");
321
322    switch (transform) {
323    case TRANSFORM_MDCT:
324        if (do_inverse) {
325            imdct_ref((float *)tab_ref, (float *)tab1, fft_nbits);
326            ff_imdct_calc(m, tab2, (float *)tab1);
327            check_diff((float *)tab_ref, tab2, fft_size, scale);
328        } else {
329            mdct_ref((float *)tab_ref, (float *)tab1, fft_nbits);
330
331            ff_mdct_calc(m, tab2, (float *)tab1);
332
333            check_diff((float *)tab_ref, tab2, fft_size / 2, scale);
334        }
335        break;
336    case TRANSFORM_FFT:
337        memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
338        ff_fft_permute(s, tab);
339        ff_fft_calc(s, tab);
340
341        fft_ref(tab_ref, tab1, fft_nbits);
342        check_diff((float *)tab_ref, (float *)tab, fft_size * 2, 1.0);
343        break;
344    case TRANSFORM_RDFT:
345        if (do_inverse) {
346            tab1[         0].im = 0;
347            tab1[fft_size_2].im = 0;
348            for (i = 1; i < fft_size_2; i++) {
349                tab1[fft_size_2+i].re =  tab1[fft_size_2-i].re;
350                tab1[fft_size_2+i].im = -tab1[fft_size_2-i].im;
351            }
352
353            memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
354            tab2[1] = tab1[fft_size_2].re;
355
356            ff_rdft_calc(r, tab2);
357            fft_ref(tab_ref, tab1, fft_nbits);
358            for (i = 0; i < fft_size; i++) {
359                tab[i].re = tab2[i];
360                tab[i].im = 0;
361            }
362            check_diff((float *)tab_ref, (float *)tab, fft_size * 2, 0.5);
363        } else {
364            for (i = 0; i < fft_size; i++) {
365                tab2[i]    = tab1[i].re;
366                tab1[i].im = 0;
367            }
368            ff_rdft_calc(r, tab2);
369            fft_ref(tab_ref, tab1, fft_nbits);
370            tab_ref[0].im = tab_ref[fft_size_2].re;
371            check_diff((float *)tab_ref, (float *)tab2, fft_size, 1.0);
372        }
373        break;
374    case TRANSFORM_DCT:
375        memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
376        ff_dct_calc(d, tab);
377        if (do_inverse) {
378            idct_ref(tab_ref, tab1, fft_nbits);
379        } else {
380            dct_ref(tab_ref, tab1, fft_nbits);
381        }
382        check_diff((float *)tab_ref, (float *)tab, fft_size, 1.0);
383        break;
384    }
385
386    /* do a speed test */
387
388    if (do_speed) {
389        int64_t time_start, duration;
390        int nb_its;
391
392        av_log(NULL, AV_LOG_INFO,"Speed test...\n");
393        /* we measure during about 1 seconds */
394        nb_its = 1;
395        for(;;) {
396            time_start = gettime();
397            for (it = 0; it < nb_its; it++) {
398                switch (transform) {
399                case TRANSFORM_MDCT:
400                    if (do_inverse) {
401                        ff_imdct_calc(m, (float *)tab, (float *)tab1);
402                    } else {
403                        ff_mdct_calc(m, (float *)tab, (float *)tab1);
404                    }
405                    break;
406                case TRANSFORM_FFT:
407                    memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
408                    ff_fft_calc(s, tab);
409                    break;
410                case TRANSFORM_RDFT:
411                    memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
412                    ff_rdft_calc(r, tab2);
413                    break;
414                case TRANSFORM_DCT:
415                    memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
416                    ff_dct_calc(d, tab2);
417                    break;
418                }
419            }
420            duration = gettime() - time_start;
421            if (duration >= 1000000)
422                break;
423            nb_its *= 2;
424        }
425        av_log(NULL, AV_LOG_INFO,"time: %0.1f us/transform [total time=%0.2f s its=%d]\n",
426               (double)duration / nb_its,
427               (double)duration / 1000000.0,
428               nb_its);
429    }
430
431    switch (transform) {
432    case TRANSFORM_MDCT:
433        ff_mdct_end(m);
434        break;
435    case TRANSFORM_FFT:
436        ff_fft_end(s);
437        break;
438    case TRANSFORM_RDFT:
439        ff_rdft_end(r);
440        break;
441    case TRANSFORM_DCT:
442        ff_dct_end(d);
443        break;
444    }
445    return 0;
446}
447