comparison mdct.c @ 7544:ee1cb5ab9f99 libavcodec

optimize imdct_half: remove tmp buffer. skip fft reinterleave pass, leaving data in a format more convenient for simd. merge post-rotate with post-reorder.

7543:f04ff5a6fb55

    double _bim = (bim);\
    (pre) = _are * _bre - _aim * _bim;\
    (pim) = _are * _bim + _aim * _bre;\
}

static void imdct_c(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
{
    int k, n4, n2, n, j;
    const uint16_t *revtab = s->fft.revtab;
    const FFTSample *tcos = s->tcos;
    const FFTSample *tsin = s->tsin;
    const FFTSample *in1, *in2;
    FFTComplex *z = (FFTComplex *)tmp;

    n = 1 << s->nbits;
    n2 = n >> 1;
    n4 = n >> 2;

    /* pre rotation */
    in1 = input;
    in2 = input + n2 - 1;
    for(k = 0; k < n4; k++) {

        in2 -= 2;
    }
    ff_fft_calc(&s->fft, z);

    /* post rotation + reordering */
    /* XXX: optimize */
    for(k = 0; k < n4; k++) {
        CMUL(z[k].re, z[k].im, z[k].re, z[k].im, tcos[k], tsin[k]);
    }
}

/**
 * Compute inverse MDCT of size N = 2^nbits

 * @param tmp N/2 samples
 */
void ff_imdct_calc(MDCTContext *s, FFTSample *output,
                   const FFTSample *input, FFTSample *tmp)
{
    int k, n8, n2, n;
    FFTComplex *z = (FFTComplex *)tmp;
    n = 1 << s->nbits;
    n2 = n >> 1;
    n8 = n >> 3;

    imdct_c(s, input, tmp);

    for(k = 0; k < n8; k++) {
        output[2*k] = -z[n8 + k].im;
        output[n2-1-2*k] = z[n8 + k].im;

        output[2*k+1] = z[n8-1-k].re;
        output[n2-1-2*k-1] = -z[n8-1-k].re;

        output[n2 + 2*k]=-z[k+n8].re;
        output[n-1- 2*k]=-z[k+n8].re;

        output[n2 + 2*k+1]=z[n8-k-1].im;
        output[n-2 - 2 * k] = z[n8-k-1].im;
    }
}

/**
 * Compute the middle half of the inverse MDCT of size N = 2^nbits,
 * thus excluding the parts that can be derived by symmetry
 * @param output N/2 samples
 * @param input N/2 samples
 * @param tmp N/2 samples
 */
void ff_imdct_half(MDCTContext *s, FFTSample *output,
                   const FFTSample *input, FFTSample *tmp)
{
    int k, n8, n4, n;
    FFTComplex *z = (FFTComplex *)tmp;
    n = 1 << s->nbits;
    n4 = n >> 2;
    n8 = n >> 3;

    imdct_c(s, input, tmp);

    for(k = 0; k < n8; k++) {
        output[n4-1-2*k]   =  z[n8+k].im;
        output[n4-1-2*k-1] = -z[n8-k-1].re;
        output[n4 + 2*k]   = -z[n8+k].re;
        output[n4 + 2*k+1] =  z[n8-k-1].im;
    }
}

/**
 * Compute MDCT of size N = 2^nbits

    int i, j, n, n8, n4, n2, n3;
    FFTSample re, im, re1, im1;
    const uint16_t *revtab = s->fft.revtab;
    const FFTSample *tcos = s->tcos;
    const FFTSample *tsin = s->tsin;
    FFTComplex *x = (FFTComplex *)tmp;

    n = 1 << s->nbits;
    n2 = n >> 1;
    n4 = n >> 2;
    n8 = n >> 3;

    }

    ff_fft_calc(&s->fft, x);

    /* post rotation */
    for(i=0;i<n4;i++) {
        re = x[i].re;
        im = x[i].im;
        CMUL(re1, im1, re, im, -tsin[i], -tcos[i]);
        out[2*i] = im1;
        out[n2-1-2*i] = re1;
    }
}

void ff_mdct_end(MDCTContext *s)
{

7544:ee1cb5ab9f99

    double _bim = (bim);\
    (pre) = _are * _bre - _aim * _bim;\
    (pim) = _are * _bim + _aim * _bre;\
}

/**
 * Compute the middle half of the inverse MDCT of size N = 2^nbits,
 * thus excluding the parts that can be derived by symmetry
 * @param output N/2 samples
 * @param input N/2 samples
 */
void ff_imdct_half(MDCTContext *s, FFTSample *output, const FFTSample *input)
{
    int k, n8, n4, n2, n, j;
    const uint16_t *revtab = s->fft.revtab;
    const FFTSample *tcos = s->tcos;
    const FFTSample *tsin = s->tsin;
    const FFTSample *in1, *in2;
    FFTComplex *z = (FFTComplex *)output;

    n = 1 << s->nbits;
    n2 = n >> 1;
    n4 = n >> 2;
    n8 = n >> 3;

    /* pre rotation */
    in1 = input;
    in2 = input + n2 - 1;
    for(k = 0; k < n4; k++) {

        in2 -= 2;
    }
    ff_fft_calc(&s->fft, z);

    /* post rotation + reordering */
    output += n4;
    for(k = 0; k < n8; k++) {
        FFTSample r0, i0, r1, i1;
        CMUL(r0, i1, z[n8-k-1].im, z[n8-k-1].re, tsin[n8-k-1], tcos[n8-k-1]);
        CMUL(r1, i0, z[n8+k  ].im, z[n8+k  ].re, tsin[n8+k  ], tcos[n8+k  ]);
        z[n8-k-1].re = r0;
        z[n8-k-1].im = i0;
        z[n8+k  ].re = r1;
        z[n8+k  ].im = i1;
    }
}

/**
 * Compute inverse MDCT of size N = 2^nbits

 * @param tmp N/2 samples
 */
void ff_imdct_calc(MDCTContext *s, FFTSample *output,
                   const FFTSample *input, FFTSample *tmp)
{
    int k;
    int n = 1 << s->nbits;
    int n2 = n >> 1;
    int n4 = n >> 2;

    ff_imdct_half(s, output+n4, input);

    for(k = 0; k < n4; k++) {
        output[k] = -output[n2-k-1];
        output[n-k-1] = output[n2+k];
    }
}

/**
 * Compute MDCT of size N = 2^nbits

    int i, j, n, n8, n4, n2, n3;
    FFTSample re, im, re1, im1;
    const uint16_t *revtab = s->fft.revtab;
    const FFTSample *tcos = s->tcos;
    const FFTSample *tsin = s->tsin;
    FFTComplex *x = (FFTComplex *)out;

    n = 1 << s->nbits;
    n2 = n >> 1;
    n4 = n >> 2;
    n8 = n >> 3;

    }

    ff_fft_calc(&s->fft, x);

    /* post rotation */
    for(i=0;i<n8;i++) {
        FFTSample r0, i0, r1, i1;
        CMUL(i1, r0, x[n8-i-1].re, x[n8-i-1].im, -tsin[n8-i-1], -tcos[n8-i-1]);
        CMUL(i0, r1, x[n8+i  ].re, x[n8+i  ].im, -tsin[n8+i  ], -tcos[n8+i  ]);
        x[n8-i-1].re = r0;
        x[n8-i-1].im = i0;
        x[n8+i  ].re = r1;
        x[n8+i  ].im = i1;
    }
}

void ff_mdct_end(MDCTContext *s)
{