annotate fft-test.c @ 11032:01bd040f8607 libavcodec

Unroll main loop so the edge==0 case is seperate. This allows many things to be simplified away. h264 decoder is overall 1% faster with a mbaff sample and 0.1% slower with the cathedral sample, probably because the slow loop filter code must be loaded into the code cache for each first MB of each row but isnt used for the following MBs.

/*
 * (c) 2002 Fabrice Bellard
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @file libavcodec/fft-test.c
 * FFT and MDCT tests.
 */

#include "libavutil/lfg.h"
#include "dsputil.h"
#include <math.h>
#include <unistd.h>
#include <sys/time.h>
#include <stdlib.h>
#include <string.h>

#undef exit

/* reference fft */

#define MUL16(a,b) ((a) * (b))

#define CMAC(pre, pim, are, aim, bre, bim) \
{\
   pre += (MUL16(are, bre) - MUL16(aim, bim));\
   pim += (MUL16(are, bim) + MUL16(bre, aim));\
}

FFTComplex *exptab;

static void fft_ref_init(int nbits, int inverse)
{
    int n, i;
    double c1, s1, alpha;

    n = 1 << nbits;
    exptab = av_malloc((n / 2) * sizeof(FFTComplex));

    for (i = 0; i < (n/2); i++) {
        alpha = 2 * M_PI * (float)i / (float)n;
        c1 = cos(alpha);
        s1 = sin(alpha);
        if (!inverse)
            s1 = -s1;
        exptab[i].re = c1;
        exptab[i].im = s1;
    }
}

static void fft_ref(FFTComplex *tabr, FFTComplex *tab, int nbits)
{
    int n, i, j, k, n2;
    double tmp_re, tmp_im, s, c;
    FFTComplex *q;

    n = 1 << nbits;
    n2 = n >> 1;
    for (i = 0; i < n; i++) {
        tmp_re = 0;
        tmp_im = 0;
        q = tab;
        for (j = 0; j < n; j++) {
            k = (i * j) & (n - 1);
            if (k >= n2) {
                c = -exptab[k - n2].re;
                s = -exptab[k - n2].im;
            } else {
                c = exptab[k].re;
                s = exptab[k].im;
            }
            CMAC(tmp_re, tmp_im, c, s, q->re, q->im);
            q++;
        }
        tabr[i].re = tmp_re;
        tabr[i].im = tmp_im;
    }
}

static void imdct_ref(float *out, float *in, int nbits)
{
    int n = 1<<nbits;
    int k, i, a;
    double sum, f;

    for (i = 0; i < n; i++) {
        sum = 0;
        for (k = 0; k < n/2; k++) {
            a = (2 * i + 1 + (n / 2)) * (2 * k + 1);
            f = cos(M_PI * a / (double)(2 * n));
            sum += f * in[k];
        }
        out[i] = -sum;
    }
}

/* NOTE: no normalisation by 1 / N is done */
static void mdct_ref(float *output, float *input, int nbits)
{
    int n = 1<<nbits;
    int k, i;
    double a, s;

    /* do it by hand */
    for (k = 0; k < n/2; k++) {
        s = 0;
        for (i = 0; i < n; i++) {
            a = (2*M_PI*(2*i+1+n/2)*(2*k+1) / (4 * n));
            s += input[i] * cos(a);
        }
        output[k] = s;
    }
}

static void idct_ref(float *output, float *input, int nbits)
{
    int n = 1<<nbits;
    int k, i;
    double a, s;

    /* do it by hand */
    for (i = 0; i < n; i++) {
        s = 0.5 * input[0];
        for (k = 1; k < n; k++) {
            a = M_PI*k*(i+0.5) / n;
            s += input[k] * cos(a);
        }
        output[i] = 2 * s / n;
    }
}
static void dct_ref(float *output, float *input, int nbits)
{
    int n = 1<<nbits;
    int k, i;
    double a, s;

    /* do it by hand */
    for (k = 0; k < n; k++) {
        s = 0;
        for (i = 0; i < n; i++) {
            a = M_PI*k*(i+0.5) / n;
            s += input[i] * cos(a);
        }
        output[k] = s;
    }
}


static float frandom(AVLFG *prng)
{
    return (int16_t)av_lfg_get(prng) / 32768.0;
}

static int64_t gettime(void)
{
    struct timeval tv;
    gettimeofday(&tv,NULL);
    return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
}

static void check_diff(float *tab1, float *tab2, int n, double scale)
{
    int i;
    double max= 0;
    double error= 0;

    for (i = 0; i < n; i++) {
        double e= fabsf(tab1[i] - (tab2[i] / scale));
        if (e >= 1e-3) {
            av_log(NULL, AV_LOG_ERROR, "ERROR %d: %f %f\n",
                   i, tab1[i], tab2[i]);
        }
        error+= e*e;
        if(e>max) max= e;
    }
    av_log(NULL, AV_LOG_INFO, "max:%f e:%g\n", max, sqrt(error)/n);
}


static void help(void)
{
    av_log(NULL, AV_LOG_INFO,"usage: fft-test [-h] [-s] [-i] [-n b]\n"
           "-h     print this help\n"
           "-s     speed test\n"
           "-m     (I)MDCT test\n"
           "-d     (I)DCT test\n"
           "-i     inverse transform test\n"
           "-n b   set the transform size to 2^b\n"
           "-f x   set scale factor for output data of (I)MDCT to x\n"
           );
    exit(1);
}

enum tf_transform {
    TRANSFORM_FFT,
    TRANSFORM_MDCT,
    TRANSFORM_RDFT,
    TRANSFORM_DCT,
};

int main(int argc, char **argv)
{
    FFTComplex *tab, *tab1, *tab_ref;
    FFTSample *tab2;
    int it, i, c;
    int do_speed = 0;
    enum tf_transform transform = TRANSFORM_FFT;
    int do_inverse = 0;
    FFTContext s1, *s = &s1;
    FFTContext m1, *m = &m1;
    RDFTContext r1, *r = &r1;
    DCTContext d1, *d = &d1;
    int fft_nbits, fft_size, fft_size_2;
    double scale = 1.0;
    AVLFG prng;
    av_lfg_init(&prng, 1);

    fft_nbits = 9;
    for(;;) {
        c = getopt(argc, argv, "hsimrdn:f:");
        if (c == -1)
            break;
        switch(c) {
        case 'h':
            help();
            break;
        case 's':
            do_speed = 1;
            break;
        case 'i':
            do_inverse = 1;
            break;
        case 'm':
            transform = TRANSFORM_MDCT;
            break;
        case 'r':
            transform = TRANSFORM_RDFT;
            break;
        case 'd':
            transform = TRANSFORM_DCT;
            break;
        case 'n':
            fft_nbits = atoi(optarg);
            break;
        case 'f':
            scale = atof(optarg);
            break;
        }
    }

    fft_size = 1 << fft_nbits;
    fft_size_2 = fft_size >> 1;
    tab = av_malloc(fft_size * sizeof(FFTComplex));
    tab1 = av_malloc(fft_size * sizeof(FFTComplex));
    tab_ref = av_malloc(fft_size * sizeof(FFTComplex));
    tab2 = av_malloc(fft_size * sizeof(FFTSample));

    switch (transform) {
    case TRANSFORM_MDCT:
        av_log(NULL, AV_LOG_INFO,"Scale factor is set to %f\n", scale);
        if (do_inverse)
            av_log(NULL, AV_LOG_INFO,"IMDCT");
        else
            av_log(NULL, AV_LOG_INFO,"MDCT");
        ff_mdct_init(m, fft_nbits, do_inverse, scale);
        break;
    case TRANSFORM_FFT:
        if (do_inverse)
            av_log(NULL, AV_LOG_INFO,"IFFT");
        else
            av_log(NULL, AV_LOG_INFO,"FFT");
        ff_fft_init(s, fft_nbits, do_inverse);
        fft_ref_init(fft_nbits, do_inverse);
        break;
    case TRANSFORM_RDFT:
        if (do_inverse)
            av_log(NULL, AV_LOG_INFO,"IRDFT");
        else
            av_log(NULL, AV_LOG_INFO,"RDFT");
        ff_rdft_init(r, fft_nbits, do_inverse ? IRDFT : RDFT);
        fft_ref_init(fft_nbits, do_inverse);
        break;
    case TRANSFORM_DCT:
        if (do_inverse)
            av_log(NULL, AV_LOG_INFO,"IDCT");
        else
            av_log(NULL, AV_LOG_INFO,"DCT");
        ff_dct_init(d, fft_nbits, do_inverse);
        break;
    }
    av_log(NULL, AV_LOG_INFO," %d test\n", fft_size);

    /* generate random data */

    for (i = 0; i < fft_size; i++) {
        tab1[i].re = frandom(&prng);
        tab1[i].im = frandom(&prng);
    }

    /* checking result */
    av_log(NULL, AV_LOG_INFO,"Checking...\n");

    switch (transform) {
    case TRANSFORM_MDCT:
        if (do_inverse) {
            imdct_ref((float *)tab_ref, (float *)tab1, fft_nbits);
            ff_imdct_calc(m, tab2, (float *)tab1);
            check_diff((float *)tab_ref, tab2, fft_size, scale);
        } else {
            mdct_ref((float *)tab_ref, (float *)tab1, fft_nbits);

            ff_mdct_calc(m, tab2, (float *)tab1);

            check_diff((float *)tab_ref, tab2, fft_size / 2, scale);
        }
        break;
    case TRANSFORM_FFT:
        memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
        ff_fft_permute(s, tab);
        ff_fft_calc(s, tab);

        fft_ref(tab_ref, tab1, fft_nbits);
        check_diff((float *)tab_ref, (float *)tab, fft_size * 2, 1.0);
        break;
    case TRANSFORM_RDFT:
        if (do_inverse) {
            tab1[         0].im = 0;
            tab1[fft_size_2].im = 0;
            for (i = 1; i < fft_size_2; i++) {
                tab1[fft_size_2+i].re =  tab1[fft_size_2-i].re;
                tab1[fft_size_2+i].im = -tab1[fft_size_2-i].im;
            }

            memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
            tab2[1] = tab1[fft_size_2].re;

            ff_rdft_calc(r, tab2);
            fft_ref(tab_ref, tab1, fft_nbits);
            for (i = 0; i < fft_size; i++) {
                tab[i].re = tab2[i];
                tab[i].im = 0;
            }
            check_diff((float *)tab_ref, (float *)tab, fft_size * 2, 0.5);
        } else {
            for (i = 0; i < fft_size; i++) {
                tab2[i]    = tab1[i].re;
                tab1[i].im = 0;
            }
            ff_rdft_calc(r, tab2);
            fft_ref(tab_ref, tab1, fft_nbits);
            tab_ref[0].im = tab_ref[fft_size_2].re;
            check_diff((float *)tab_ref, (float *)tab2, fft_size, 1.0);
        }
        break;
    case TRANSFORM_DCT:
        memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
        ff_dct_calc(d, tab);
        if (do_inverse) {
            idct_ref(tab_ref, tab1, fft_nbits);
        } else {
            dct_ref(tab_ref, tab1, fft_nbits);
        }
        check_diff((float *)tab_ref, (float *)tab, fft_size, 1.0);
        break;
    }

    /* do a speed test */

    if (do_speed) {
        int64_t time_start, duration;
        int nb_its;

        av_log(NULL, AV_LOG_INFO,"Speed test...\n");
        /* we measure during about 1 seconds */
        nb_its = 1;
        for(;;) {
            time_start = gettime();
            for (it = 0; it < nb_its; it++) {
                switch (transform) {
                case TRANSFORM_MDCT:
                    if (do_inverse) {
                        ff_imdct_calc(m, (float *)tab, (float *)tab1);
                    } else {
                        ff_mdct_calc(m, (float *)tab, (float *)tab1);
                    }
                    break;
                case TRANSFORM_FFT:
                    memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
                    ff_fft_calc(s, tab);
                    break;
                case TRANSFORM_RDFT:
                    memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
                    ff_rdft_calc(r, tab2);
                    break;
                case TRANSFORM_DCT:
                    memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
                    ff_dct_calc(d, tab2);
                    break;
                }
            }
            duration = gettime() - time_start;
            if (duration >= 1000000)
                break;
            nb_its *= 2;
        }
        av_log(NULL, AV_LOG_INFO,"time: %0.1f us/transform [total time=%0.2f s its=%d]\n",
               (double)duration / nb_its,
               (double)duration / 1000000.0,
               nb_its);
    }

    switch (transform) {
    case TRANSFORM_MDCT:
        ff_mdct_end(m);
        break;
    case TRANSFORM_FFT:
        ff_fft_end(s);
        break;
    case TRANSFORM_RDFT:
        ff_rdft_end(r);
        break;
    case TRANSFORM_DCT:
        ff_dct_end(d);
        break;
    }
    return 0;
}