cfft.c

这是著名的TCPMP播放器在WINDWOWS,和WINCE下编译通过的源程序.笔者对其中的LIBMAD库做了针对ARM MPU的优化. 并增加了词幕功能.

#else
                    ComplexMult(&IM(ch[ah+l1*ido]), &RE(ch[ah+l1*ido]),
                        IM(d2), RE(d2), RE(wa1[i]), IM(wa1[i]));
                    ComplexMult(&IM(ch[ah+2*l1*ido]), &RE(ch[ah+2*l1*ido]),
                        IM(d3), RE(d3), RE(wa2[i]), IM(wa2[i]));
                    ComplexMult(&IM(ch[ah+3*l1*ido]), &RE(ch[ah+3*l1*ido]),
                        IM(d4), RE(d4), RE(wa3[i]), IM(wa3[i]));
                    ComplexMult(&IM(ch[ah+4*l1*ido]), &RE(ch[ah+4*l1*ido]),
                        IM(d5), RE(d5), RE(wa4[i]), IM(wa4[i]));
#endif
                }
            }
        }
    }
}


/*----------------------------------------------------------------------
   cfftf1, cfftf, cfftb, cffti1, cffti. Complex FFTs.
  ----------------------------------------------------------------------*/

static INLINE void cfftf1pos(uint16_t n, complex_t *c, complex_t *ch,
                             const uint16_t *ifac, const complex_t *wa,
                             const int8_t isign)
{
    uint16_t i;
    uint16_t k1, l1, l2;
    uint16_t na, nf, ip, iw, ix2, ix3, ix4, ido, idl1;

    nf = ifac[1];
    na = 0;
    l1 = 1;
    iw = 0;

    for (k1 = 2; k1 <= nf+1; k1++)
    {
        ip = ifac[k1];
        l2 = ip*l1;
        ido = n / l2;
        idl1 = ido*l1;

        switch (ip)
        {
        case 4:
            ix2 = iw + ido;
            ix3 = ix2 + ido;

            if (na == 0)
                passf4pos((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)c, ch, &wa[iw], &wa[ix2], &wa[ix3]);
            else
                passf4pos((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)ch, c, &wa[iw], &wa[ix2], &wa[ix3]);

            na = 1 - na;
            break;
        case 2:
            if (na == 0)
                passf2pos((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)c, ch, &wa[iw]);
            else
                passf2pos((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)ch, c, &wa[iw]);

            na = 1 - na;
            break;
        case 3:
            ix2 = iw + ido;

            if (na == 0)
                passf3((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)c, ch, &wa[iw], &wa[ix2], isign);
            else
                passf3((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)ch, c, &wa[iw], &wa[ix2], isign);

            na = 1 - na;
            break;
        case 5:
            ix2 = iw + ido;
            ix3 = ix2 + ido;
            ix4 = ix3 + ido;

            if (na == 0)
                passf5((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)c, ch, &wa[iw], &wa[ix2], &wa[ix3], &wa[ix4], isign);
            else
                passf5((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)ch, c, &wa[iw], &wa[ix2], &wa[ix3], &wa[ix4], isign);

            na = 1 - na;
            break;
        }

        l1 = l2;
        iw += (ip-1) * ido;
    }

    if (na == 0)
        return;

    for (i = 0; i < n; i++)
    {
        RE(c[i]) = RE(ch[i]);
        IM(c[i]) = IM(ch[i]);
    }
}

static INLINE void cfftf1neg(uint16_t n, complex_t *c, complex_t *ch,
                             const uint16_t *ifac, const complex_t *wa,
                             const int8_t isign)
{
    uint16_t i;
    uint16_t k1, l1, l2;
    uint16_t na, nf, ip, iw, ix2, ix3, ix4, ido, idl1;

    nf = ifac[1];
    na = 0;
    l1 = 1;
    iw = 0;

    for (k1 = 2; k1 <= nf+1; k1++)
    {
        ip = ifac[k1];
        l2 = ip*l1;
        ido = n / l2;
        idl1 = ido*l1;

        switch (ip)
        {
        case 4:
            ix2 = iw + ido;
            ix3 = ix2 + ido;

            if (na == 0)
                passf4neg((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)c, ch, &wa[iw], &wa[ix2], &wa[ix3]);
            else
                passf4neg((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)ch, c, &wa[iw], &wa[ix2], &wa[ix3]);

            na = 1 - na;
            break;
        case 2:
            if (na == 0)
                passf2neg((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)c, ch, &wa[iw]);
            else
                passf2neg((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)ch, c, &wa[iw]);

            na = 1 - na;
            break;
        case 3:
            ix2 = iw + ido;

            if (na == 0)
                passf3((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)c, ch, &wa[iw], &wa[ix2], isign);
            else
                passf3((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)ch, c, &wa[iw], &wa[ix2], isign);

            na = 1 - na;
            break;
        case 5:
            ix2 = iw + ido;
            ix3 = ix2 + ido;
            ix4 = ix3 + ido;

            if (na == 0)
                passf5((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)c, ch, &wa[iw], &wa[ix2], &wa[ix3], &wa[ix4], isign);
            else
                passf5((const uint16_t)ido, (const uint16_t)l1, (const complex_t*)ch, c, &wa[iw], &wa[ix2], &wa[ix3], &wa[ix4], isign);

            na = 1 - na;
            break;
        }

        l1 = l2;
        iw += (ip-1) * ido;
    }

    if (na == 0)
        return;

    for (i = 0; i < n; i++)
    {
        RE(c[i]) = RE(ch[i]);
        IM(c[i]) = IM(ch[i]);
    }
}

void cfftf(cfft_info *cfft, complex_t *c)
{
    cfftf1neg(cfft->n, c, cfft->work, (const uint16_t*)cfft->ifac, (const complex_t*)cfft->tab, -1);
}

void cfftb(cfft_info *cfft, complex_t *c)
{
    cfftf1pos(cfft->n, c, cfft->work, (const uint16_t*)cfft->ifac, (const complex_t*)cfft->tab, +1);
}

static void cffti1(uint16_t n, complex_t *wa, uint16_t *ifac)
{
    static const uint16_t ntryh[4] = {3, 4, 2, 5};
#ifndef FIXED_POINT
    real_t arg, argh, argld, fi;
    uint16_t ido, ipm;
    uint16_t i1, k1, l1, l2;
    uint16_t ld, ii, ip;
#endif
    uint16_t ntry = 0, i, j;
    uint16_t ib;
    uint16_t nf, nl, nq, nr;

    nl = n;
    nf = 0;
    j = 0;

startloop:
    j++;

    if (j <= 4)
        ntry = ntryh[j-1];
    else
        ntry += 2;

    do
    {
        nq = nl / ntry;
        nr = nl - ntry*nq;

        if (nr != 0)
            goto startloop;

        nf++;
        ifac[nf+1] = ntry;
        nl = nq;

        if (ntry == 2 && nf != 1)
        {
            for (i = 2; i <= nf; i++)
            {
                ib = nf - i + 2;
                ifac[ib+1] = ifac[ib];
            }
            ifac[2] = 2;
        }
    } while (nl != 1);

    ifac[0] = n;
    ifac[1] = nf;

#ifndef FIXED_POINT
    argh = (real_t)2.0*(real_t)M_PI / (real_t)n;
    i = 0;
    l1 = 1;

    for (k1 = 1; k1 <= nf; k1++)
    {
        ip = ifac[k1+1];
        ld = 0;
        l2 = l1*ip;
        ido = n / l2;
        ipm = ip - 1;

        for (j = 0; j < ipm; j++)
        {
            i1 = i;
            RE(wa[i]) = 1.0;
            IM(wa[i]) = 0.0;
            ld += l1;
            fi = 0;
            argld = ld*argh;

            for (ii = 0; ii < ido; ii++)
            {
                i++;
                fi++;
                arg = fi * argld;
                RE(wa[i]) = (real_t)cos(arg);
#if 1
                IM(wa[i]) = (real_t)sin(arg);
#else
                IM(wa[i]) = (real_t)-sin(arg);
#endif
            }

            if (ip > 5)
            {
                RE(wa[i1]) = RE(wa[i]);
                IM(wa[i1]) = IM(wa[i]);
            }
        }
        l1 = l2;
    }
#endif
}

cfft_info *cffti(uint16_t n)
{
    cfft_info *cfft = (cfft_info*)faad_malloc(sizeof(cfft_info));

    cfft->n = n;
    cfft->work = (complex_t*)faad_malloc(n*sizeof(complex_t));

#ifndef FIXED_POINT
    cfft->tab = (complex_t*)faad_malloc(n*sizeof(complex_t));

    cffti1(n, cfft->tab, cfft->ifac);
#else
    cffti1(n, NULL, cfft->ifac);

    switch (n)
    {
    case 64: cfft->tab = (complex_t*)cfft_tab_64; break;
    case 512: cfft->tab = (complex_t*)cfft_tab_512; break;
#ifdef LD_DEC
    case 256: cfft->tab = (complex_t*)cfft_tab_256; break;
#endif

#ifdef ALLOW_SMALL_FRAMELENGTH
    case 60: cfft->tab = (complex_t*)cfft_tab_60; break;
    case 480: cfft->tab = (complex_t*)cfft_tab_480; break;
#ifdef LD_DEC
    case 240: cfft->tab = (complex_t*)cfft_tab_240; break;
#endif
#endif
    case 128: cfft->tab = (complex_t*)cfft_tab_128; break;
    }
#endif

    return cfft;
}

void cfftu(cfft_info *cfft)
{
    if (cfft->work) faad_free(cfft->work);
#ifndef FIXED_POINT
    if (cfft->tab) faad_free(cfft->tab);
#endif

    if (cfft) faad_free(cfft);
}

#endif