Fixed

/* 1, 2, 4, 8, ..., 2^17 */
static const uint16_t invmod2vals[18] = { 0x0, 0x0, 0x3, 0x7, 0xf, 0xf, 0x2f, 0x6f, 0xef, 0xef, 0x2ef, 0x6ef, 0xeef, 0xeef, 0x2eef, 0x6eef, 0xeeef, 0xeeef };

void ff_fft15_transform(FFTContext *s, FFTComplex *out, FFTComplex *in)
{
    FFTComplex fft15in[15];
    unsigned int unused, nbits = s->nbits, n = (1U << nbits), Ntot = 15*n;
    unsigned int inv_1 = (n << ((4 - nbits)&3));
    unsigned int inv_2 = invmod2vals[nbits];
    unsigned int i, j;

    for (i = 0; i < n; i++) {
        unsigned int l = s->revtab_fwd[i];
        for (j = 0; j < 15; j++) {
            unsigned int q = (n*j)/15;
            q += i;
            q >>= nbits;
            k = 15*i + ((j - q*15)<<nbits);
            fft15in[j] = in[k];
        }
        fft15(out+l, fft15in, n);
    }
    for (i = 0; i < 15; i++) {
        fft_calc_internal(out+n*i, nbits); // this is "ff_fft_calc_interleave_sse" in the ffmpeg codebase
    }
    memcpy(in, out, Ntot * sizeof(FFTComplex));
    for (j = 0; j < 15; j++) {
        for (i = 0; i < n; i++) {
            unsigned int q = (j*inv_1)/15;
            q += (i*inv_2);
            q >>= nbits;
            k = (i*inv_2*15 + j*inv_1) - q*Ntot;
            out[k] = in[n*j + i];
        }
    }
}