Ejemplo n.º 1
0
/**
 * Calculate rate distortion cost for quantizing with given codebook
 *
 * @return quantization distortion
 */
static float quantize_and_encode_band_cost(struct AACEncContext *s,
                                PutBitContext *pb, const float *in,
                                const float *scaled, int size, int scale_idx,
                                int cb, const float lambda, const float uplim,
                                int *bits)
{
    const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
    const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
    const float CLIPPED_ESCAPE = 165140.0f*IQ;
    int i, j, k;
    float cost = 0;
    const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
    int resbits = 0;
#ifndef USE_REALLY_FULL_SEARCH
    const float  Q34 = sqrtf(Q * sqrtf(Q));
    const int range  = aac_cb_range[cb];
    const int maxval = aac_cb_maxval[cb];
    int offs[4];
#endif /* USE_REALLY_FULL_SEARCH */

    if (!cb) {
        for (i = 0; i < size; i++)
            cost += in[i]*in[i];
        if (bits)
            *bits = 0;
        return cost * lambda;
    }
#ifndef USE_REALLY_FULL_SEARCH
    offs[0] = 1;
    for (i = 1; i < dim; i++)
        offs[i] = offs[i-1]*range;
    if (!scaled) {
        abs_pow34_v(s->scoefs, in, size);
        scaled = s->scoefs;
    }
    quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
#endif /* USE_REALLY_FULL_SEARCH */
    for (i = 0; i < size; i += dim) {
        float mincost;
        int minidx  = 0;
        int minbits = 0;
        const float *vec;
#ifndef USE_REALLY_FULL_SEARCH
        int (*quants)[2] = &s->qcoefs[i];
        mincost = 0.0f;
        for (j = 0; j < dim; j++)
            mincost += in[i+j]*in[i+j];
        minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
        minbits = ff_aac_spectral_bits[cb-1][minidx];
        mincost = mincost * lambda + minbits;
        for (j = 0; j < (1<<dim); j++) {
            float rd = 0.0f;
            int curbits;
            int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
            int same   = 0;
            for (k = 0; k < dim; k++) {
                if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
                    same = 1;
                    break;
                }
            }
            if (same)
                continue;
            for (k = 0; k < dim; k++)
                curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
            curbits =  ff_aac_spectral_bits[cb-1][curidx];
            vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
#else
        mincost = INFINITY;
        vec = ff_aac_codebook_vectors[cb-1];
        for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
            float rd = 0.0f;
            int curbits = ff_aac_spectral_bits[cb-1][j];
            int curidx = j;
#endif /* USE_REALLY_FULL_SEARCH */
            if (IS_CODEBOOK_UNSIGNED(cb)) {
                for (k = 0; k < dim; k++) {
                    float t = fabsf(in[i+k]);
                    float di;
                    if (vec[k] == 64.0f) { //FIXME: slow
                        //do not code with escape sequence small values
                        if (t < 39.0f*IQ) {
                            rd = INFINITY;
                            break;
                        }
                        if (t >= CLIPPED_ESCAPE) {
                            di = t - CLIPPED_ESCAPE;
                            curbits += 21;
                        } else {
                            int c = av_clip(quant(t, Q), 0, 8191);
                            di = t - c*cbrtf(c)*IQ;
                            curbits += av_log2(c)*2 - 4 + 1;
                        }
                    } else {
                        di = t - vec[k]*IQ;
                    }
                    if (vec[k] != 0.0f)
                        curbits++;
                    rd += di*di;
                }
            } else {
                for (k = 0; k < dim; k++) {
                    float di = in[i+k] - vec[k]*IQ;
                    rd += di*di;
                }
            }
            rd = rd * lambda + curbits;
            if (rd < mincost) {
                mincost = rd;
                minidx  = curidx;
                minbits = curbits;
            }
        }
        cost    += mincost;
        resbits += minbits;
        if (cost >= uplim)
            return uplim;
        if (pb) {
        put_bits(pb, ff_aac_spectral_bits[cb-1][minidx], ff_aac_spectral_codes[cb-1][minidx]);
        if (IS_CODEBOOK_UNSIGNED(cb))
            for (j = 0; j < dim; j++)
                if (ff_aac_codebook_vectors[cb-1][minidx*dim+j] != 0.0f)
                    put_bits(pb, 1, in[i+j] < 0.0f);
        if (cb == ESC_BT) {
            for (j = 0; j < 2; j++) {
                if (ff_aac_codebook_vectors[cb-1][minidx*2+j] == 64.0f) {
                    int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
                    int len = av_log2(coef);

                    put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
                    put_bits(pb, len, coef & ((1 << len) - 1));
                }
            }
        }
        }
    }

    if (bits)
        *bits = resbits;
    return cost;
}
static float quantize_band_cost(struct AACEncContext *s, const float *in,
                                const float *scaled, int size, int scale_idx,
                                int cb, const float lambda, const float uplim,
                                int *bits)
{
    return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
                                         cb, lambda, uplim, bits);
}

static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
                                     const float *in, int size, int scale_idx,
                                     int cb, const float lambda)
{
    quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
                                  INFINITY, NULL);
}
Ejemplo n.º 2
0
/**
 * Calculate rate distortion cost for quantizing with given codebook
 *
 * @return quantization distortion
 */
static av_always_inline float quantize_and_encode_band_cost_template(
                                struct AACEncContext *s,
                                PutBitContext *pb, const float *in,
                                const float *scaled, int size, int scale_idx,
                                int cb, const float lambda, const float uplim,
                                int *bits, int BT_ZERO, int BT_UNSIGNED,
                                int BT_PAIR, int BT_ESC)
{
    const float IQ = ff_aac_pow2sf_tab[POW_SF2_ZERO + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
    const float  Q = ff_aac_pow2sf_tab[POW_SF2_ZERO - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
    const float CLIPPED_ESCAPE = 165140.0f*IQ;
    int i, j;
    float cost = 0;
    const int dim = BT_PAIR ? 2 : 4;
    int resbits = 0;
    const float  Q34 = sqrtf(Q * sqrtf(Q));
    const int range  = aac_cb_range[cb];
    const int maxval = aac_cb_maxval[cb];
    int off;

    if (BT_ZERO) {
        for (i = 0; i < size; i++)
            cost += in[i]*in[i];
        if (bits)
            *bits = 0;
        return cost * lambda;
    }
    if (!scaled) {
        abs_pow34_v(s->scoefs, in, size);
        scaled = s->scoefs;
    }
    quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, maxval);
    if (BT_UNSIGNED) {
        off = 0;
    } else {
        off = maxval;
    }
    for (i = 0; i < size; i += dim) {
        const float *vec;
        int *quants = s->qcoefs + i;
        int curidx = 0;
        int curbits;
        float rd = 0.0f;
        for (j = 0; j < dim; j++) {
            curidx *= range;
            curidx += quants[j] + off;
        }
        curbits =  ff_aac_spectral_bits[cb-1][curidx];
        vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
        if (BT_UNSIGNED) {
            for (j = 0; j < dim; j++) {
                float t = fabsf(in[i+j]);
                float di;
                if (BT_ESC && vec[j] == 64.0f) { //FIXME: slow
                    if (t >= CLIPPED_ESCAPE) {
                        di = t - CLIPPED_ESCAPE;
                        curbits += 21;
                    } else {
                        int c = av_clip(quant(t, Q), 0, 8191);
                        di = t - c*cbrtf(c)*IQ;
                        curbits += av_log2(c)*2 - 4 + 1;
                    }
                } else {
                    di = t - vec[j]*IQ;
                }
                if (vec[j] != 0.0f)
                    curbits++;
                rd += di*di;
            }
        } else {
            for (j = 0; j < dim; j++) {
                float di = in[i+j] - vec[j]*IQ;
                rd += di*di;
            }
        }
        cost    += rd * lambda + curbits;
        resbits += curbits;
        if (cost >= uplim)
            return uplim;
        if (pb) {
            put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
            if (BT_UNSIGNED)
                for (j = 0; j < dim; j++)
                    if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
                        put_bits(pb, 1, in[i+j] < 0.0f);
            if (BT_ESC) {
                for (j = 0; j < 2; j++) {
                    if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
                        int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
                        int len = av_log2(coef);

                        put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
                        put_bits(pb, len, coef & ((1 << len) - 1));
                    }
                }
            }
        }
    }

    if (bits)
        *bits = resbits;
    return cost;
}
Ejemplo n.º 3
0
Archivo: aaccoder.c Proyecto: 119/ipnc
/**
 * Calculate rate distortion cost for quantizing with given codebook
 *
 * @return quantization distortion
 */
static float quantize_band_cost(struct AACEncContext *s, const float *in,
                                const float *scaled, int size, int scale_idx,
                                int cb, const float lambda, const float uplim,
                                int *bits)
{
    const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
    const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
    const float CLIPPED_ESCAPE = 165140.0f*IQ;
    int i, j, k;
    float cost = 0;
    const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
    int resbits = 0;
#ifndef USE_REALLY_FULL_SEARCH
    const float  Q34 = sqrtf(Q * sqrtf(Q));
    const int range  = aac_cb_range[cb];
    const int maxval = aac_cb_maxval[cb];
    int offs[4];
#endif /* USE_REALLY_FULL_SEARCH */

    if (!cb) {
        for (i = 0; i < size; i++)
            cost += in[i]*in[i];
        if (bits)
            *bits = 0;
        return cost * lambda;
    }
#ifndef USE_REALLY_FULL_SEARCH
    offs[0] = 1;
    for (i = 1; i < dim; i++)
        offs[i] = offs[i-1]*range;
    quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
#endif /* USE_REALLY_FULL_SEARCH */
    for (i = 0; i < size; i += dim) {
        float mincost;
        int minidx  = 0;
        int minbits = 0;
        const float *vec;
#ifndef USE_REALLY_FULL_SEARCH
        int (*quants)[2] = &s->qcoefs[i];
        mincost = 0.0f;
        for (j = 0; j < dim; j++)
            mincost += in[i+j]*in[i+j];
        minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
        minbits = ff_aac_spectral_bits[cb-1][minidx];
        mincost = mincost * lambda + minbits;
        for (j = 0; j < (1<<dim); j++) {
            float rd = 0.0f;
            int curbits;
            int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
            int same   = 0;
            for (k = 0; k < dim; k++) {
                if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
                    same = 1;
                    break;
                }
            }
            if (same)
                continue;
            for (k = 0; k < dim; k++)
                curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
            curbits =  ff_aac_spectral_bits[cb-1][curidx];
            vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
#else
        mincost = INFINITY;
        vec = ff_aac_codebook_vectors[cb-1];
        for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
            float rd = 0.0f;
            int curbits = ff_aac_spectral_bits[cb-1][j];
#endif /* USE_REALLY_FULL_SEARCH */
            if (IS_CODEBOOK_UNSIGNED(cb)) {
                for (k = 0; k < dim; k++) {
                    float t = fabsf(in[i+k]);
                    float di;
                    if (vec[k] == 64.0f) { //FIXME: slow
                        //do not code with escape sequence small values
                        if (t < 39.0f*IQ) {
                            rd = INFINITY;
                            break;
                        }
                        if (t >= CLIPPED_ESCAPE) {
                            di = t - CLIPPED_ESCAPE;
                            curbits += 21;
                        } else {
                            int c = av_clip(quant(t, Q), 0, 8191);
                            di = t - c*cbrtf(c)*IQ;
                            curbits += av_log2(c)*2 - 4 + 1;
                        }
                    } else {
                        di = t - vec[k]*IQ;
                    }
                    if (vec[k] != 0.0f)
                        curbits++;
                    rd += di*di;
                }
            } else {
                for (k = 0; k < dim; k++) {
                    float di = in[i+k] - vec[k]*IQ;
                    rd += di*di;
                }
            }
            rd = rd * lambda + curbits;
            if (rd < mincost) {
                mincost = rd;
                minidx  = j;
                minbits = curbits;
            }
        }
        cost    += mincost;
        resbits += minbits;
        if (cost >= uplim)
            return uplim;
    }

    if (bits)
        *bits = resbits;
    return cost;
}

static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
                                     const float *in, int size, int scale_idx,
                                     int cb, const float lambda)
{
    const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
    const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
    const float CLIPPED_ESCAPE = 165140.0f*IQ;
    const int dim = (cb < FIRST_PAIR_BT) ? 4 : 2;
    int i, j, k;
#ifndef USE_REALLY_FULL_SEARCH
    const float  Q34 = sqrtf(Q * sqrtf(Q));
    const int range  = aac_cb_range[cb];
    const int maxval = aac_cb_maxval[cb];
    int offs[4];
    float *scaled = s->scoefs;
#endif /* USE_REALLY_FULL_SEARCH */

//START_TIMER
    if (!cb)
        return;

#ifndef USE_REALLY_FULL_SEARCH
    offs[0] = 1;
    for (i = 1; i < dim; i++)
        offs[i] = offs[i-1]*range;
    abs_pow34_v(scaled, in, size);
    quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
#endif /* USE_REALLY_FULL_SEARCH */
    for (i = 0; i < size; i += dim) {
        float mincost;
        int minidx  = 0;
        int minbits = 0;
        const float *vec;
#ifndef USE_REALLY_FULL_SEARCH
        int (*quants)[2] = &s->qcoefs[i];
        mincost = 0.0f;
        for (j = 0; j < dim; j++)
            mincost += in[i+j]*in[i+j];
        minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
        minbits = ff_aac_spectral_bits[cb-1][minidx];
        mincost = mincost * lambda + minbits;
        for (j = 0; j < (1<<dim); j++) {
            float rd = 0.0f;
            int curbits;
            int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
            int same   = 0;
            for (k = 0; k < dim; k++) {
                if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
                    same = 1;
                    break;
                }
            }
            if (same)
                continue;
            for (k = 0; k < dim; k++)
                curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
            curbits =  ff_aac_spectral_bits[cb-1][curidx];
            vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
#else
        vec = ff_aac_codebook_vectors[cb-1];
        mincost = INFINITY;
        for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
            float rd = 0.0f;
            int curbits = ff_aac_spectral_bits[cb-1][j];
            int curidx  = j;
#endif /* USE_REALLY_FULL_SEARCH */
            if (IS_CODEBOOK_UNSIGNED(cb)) {
                for (k = 0; k < dim; k++) {
                    float t = fabsf(in[i+k]);
                    float di;
                    if (vec[k] == 64.0f) { //FIXME: slow
                        //do not code with escape sequence small values
                        if (t < 39.0f*IQ) {
                            rd = INFINITY;
                            break;
                        }
                        if (t >= CLIPPED_ESCAPE) {
                            di = t - CLIPPED_ESCAPE;
                            curbits += 21;
                        } else {
                            int c = av_clip(quant(t, Q), 0, 8191);
                            di = t - c*cbrtf(c)*IQ;
                            curbits += av_log2(c)*2 - 4 + 1;
                        }
                    } else {
                        di = t - vec[k]*IQ;
                    }
                    if (vec[k] != 0.0f)
                        curbits++;
                    rd += di*di;
                }
            } else {
                for (k = 0; k < dim; k++) {
                    float di = in[i+k] - vec[k]*IQ;
                    rd += di*di;
                }
            }
            rd = rd * lambda + curbits;
            if (rd < mincost) {
                mincost = rd;
                minidx  = curidx;
                minbits = curbits;
            }
        }
        put_bits(pb, ff_aac_spectral_bits[cb-1][minidx], ff_aac_spectral_codes[cb-1][minidx]);
        if (IS_CODEBOOK_UNSIGNED(cb))
            for (j = 0; j < dim; j++)
                if (ff_aac_codebook_vectors[cb-1][minidx*dim+j] != 0.0f)
                    put_bits(pb, 1, in[i+j] < 0.0f);
        if (cb == ESC_BT) {
            for (j = 0; j < 2; j++) {
                if (ff_aac_codebook_vectors[cb-1][minidx*2+j] == 64.0f) {
                    int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
                    int len = av_log2(coef);

                    put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
                    put_bits(pb, len, coef & ((1 << len) - 1));
                }
            }
        }
    }
//STOP_TIMER("quantize_and_encode")
}

/**
 * structure used in optimal codebook search
 */
typedef struct BandCodingPath {
    int prev_idx; ///< pointer to the previous path point
    float cost;   ///< path cost
    int run;
} BandCodingPath;

/**
 * Encode band info for single window group bands.
 */
static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
                                     int win, int group_len, const float lambda)
{
    BandCodingPath path[120][12];
    int w, swb, cb, start, start2, size;
    int i, j;
    const int max_sfb  = sce->ics.max_sfb;
    const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
    const int run_esc  = (1 << run_bits) - 1;
    int idx, ppos, count;
    int stackrun[120], stackcb[120], stack_len;
    float next_minrd = INFINITY;
    int next_mincb = 0;

    abs_pow34_v(s->scoefs, sce->coeffs, 1024);
    start = win*128;
    for (cb = 0; cb < 12; cb++) {
        path[0][cb].cost     = 0.0f;
        path[0][cb].prev_idx = -1;
        path[0][cb].run      = 0;
    }
    for (swb = 0; swb < max_sfb; swb++) {
        start2 = start;
        size = sce->ics.swb_sizes[swb];
        if (sce->zeroes[win*16 + swb]) {
            for (cb = 0; cb < 12; cb++) {
                path[swb+1][cb].prev_idx = cb;
                path[swb+1][cb].cost     = path[swb][cb].cost;
                path[swb+1][cb].run      = path[swb][cb].run + 1;
            }
        } else {
            float minrd = next_minrd;
            int mincb = next_mincb;
            next_minrd = INFINITY;
            next_mincb = 0;
            for (cb = 0; cb < 12; cb++) {
                float cost_stay_here, cost_get_here;
                float rd = 0.0f;
                for (w = 0; w < group_len; w++) {
                    FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
                    rd += quantize_band_cost(s, sce->coeffs + start + w*128,
                                             s->scoefs + start + w*128, size,
                                             sce->sf_idx[(win+w)*16+swb], cb,
                                             lambda / band->threshold, INFINITY, NULL);
                }
                cost_stay_here = path[swb][cb].cost + rd;
                cost_get_here  = minrd              + rd + run_bits + 4;
                if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
                    != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
                    cost_stay_here += run_bits;
                if (cost_get_here < cost_stay_here) {
                    path[swb+1][cb].prev_idx = mincb;
                    path[swb+1][cb].cost     = cost_get_here;
                    path[swb+1][cb].run      = 1;
                } else {
                    path[swb+1][cb].prev_idx = cb;
                    path[swb+1][cb].cost     = cost_stay_here;
                    path[swb+1][cb].run      = path[swb][cb].run + 1;
                }
                if (path[swb+1][cb].cost < next_minrd) {
                    next_minrd = path[swb+1][cb].cost;
                    next_mincb = cb;
                }
            }
        }
        start += sce->ics.swb_sizes[swb];
    }

    //convert resulting path from backward-linked list
    stack_len = 0;
    idx       = 0;
    for (cb = 1; cb < 12; cb++)
        if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
            idx = cb;
    ppos = max_sfb;
    while (ppos > 0) {
        cb = idx;
        stackrun[stack_len] = path[ppos][cb].run;
        stackcb [stack_len] = cb;
        idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
        ppos -= path[ppos][cb].run;
        stack_len++;
    }
    //perform actual band info encoding
    start = 0;
    for (i = stack_len - 1; i >= 0; i--) {
        put_bits(&s->pb, 4, stackcb[i]);
        count = stackrun[i];
        memset(sce->zeroes + win*16 + start, !stackcb[i], count);
        //XXX: memset when band_type is also uint8_t
        for (j = 0; j < count; j++) {
            sce->band_type[win*16 + start] =  stackcb[i];
            start++;
        }
        while (count >= run_esc) {
            put_bits(&s->pb, run_bits, run_esc);
            count -= run_esc;
        }
        put_bits(&s->pb, run_bits, count);
    }
}

typedef struct TrellisPath {
    float cost;
    int prev;
    int min_val;
    int max_val;
} TrellisPath;

#define TRELLIS_STAGES 121
#define TRELLIS_STATES 256

static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
                                       SingleChannelElement *sce,
                                       const float lambda)
{
    int q, w, w2, g, start = 0;
    int i, j;
    int idx;
    TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
    int bandaddr[TRELLIS_STAGES];
    int minq;
    float mincost;

    for (i = 0; i < TRELLIS_STATES; i++) {
        paths[0][i].cost    = 0.0f;
        paths[0][i].prev    = -1;
        paths[0][i].min_val = i;
        paths[0][i].max_val = i;
    }
    for (j = 1; j < TRELLIS_STAGES; j++) {
        for (i = 0; i < TRELLIS_STATES; i++) {
            paths[j][i].cost    = INFINITY;
            paths[j][i].prev    = -2;
            paths[j][i].min_val = INT_MAX;
            paths[j][i].max_val = 0;
        }
    }
    idx = 1;
    abs_pow34_v(s->scoefs, sce->coeffs, 1024);
    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
        start = w*128;
        for (g = 0; g < sce->ics.num_swb; g++) {
            const float *coefs = sce->coeffs + start;
            float qmin, qmax;
            int nz = 0;

            bandaddr[idx] = w * 16 + g;
            qmin = INT_MAX;
            qmax = 0.0f;
            for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
                if (band->energy <= band->threshold || band->threshold == 0.0f) {
                    sce->zeroes[(w+w2)*16+g] = 1;
                    continue;
                }
                sce->zeroes[(w+w2)*16+g] = 0;
                nz = 1;
                for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
                    float t = fabsf(coefs[w2*128+i]);
                    if (t > 0.0f)
                        qmin = FFMIN(qmin, t);
                    qmax = FFMAX(qmax, t);
                }
            }
            if (nz) {
                int minscale, maxscale;
                float minrd = INFINITY;
                //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
                minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
                //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
                maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
                for (q = minscale; q < maxscale; q++) {
                    float dists[12], dist;
                    memset(dists, 0, sizeof(dists));
                    for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                        FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
                        int cb;
                        for (cb = 0; cb <= ESC_BT; cb++)
                            dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
                                                            q, cb, lambda / band->threshold, INFINITY, NULL);
                    }
                    dist = dists[0];
                    for (i = 1; i <= ESC_BT; i++)
                        dist = FFMIN(dist, dists[i]);
                    minrd = FFMIN(minrd, dist);

                    for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, TRELLIS_STATES); i++) {
                        float cost;
                        int minv, maxv;
                        if (isinf(paths[idx - 1][i].cost))
                            continue;
                        cost = paths[idx - 1][i].cost + dist
                               + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
                        minv = FFMIN(paths[idx - 1][i].min_val, q);
                        maxv = FFMAX(paths[idx - 1][i].max_val, q);
                        if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
                            paths[idx][q].cost    = cost;
                            paths[idx][q].prev    = i;
                            paths[idx][q].min_val = minv;
                            paths[idx][q].max_val = maxv;
                        }
                    }
                }
            } else {
                for (q = 0; q < TRELLIS_STATES; q++) {
                    if (!isinf(paths[idx - 1][q].cost)) {
                        paths[idx][q].cost = paths[idx - 1][q].cost + 1;
                        paths[idx][q].prev = q;
                        paths[idx][q].min_val = FFMIN(paths[idx - 1][q].min_val, q);
                        paths[idx][q].max_val = FFMAX(paths[idx - 1][q].max_val, q);
                        continue;
                    }
                    for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, TRELLIS_STATES); i++) {
                        float cost;
                        int minv, maxv;
                        if (isinf(paths[idx - 1][i].cost))
                            continue;
                        cost = paths[idx - 1][i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
                        minv = FFMIN(paths[idx - 1][i].min_val, q);
                        maxv = FFMAX(paths[idx - 1][i].max_val, q);
                        if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
                            paths[idx][q].cost    = cost;
                            paths[idx][q].prev    = i;
                            paths[idx][q].min_val = minv;
                            paths[idx][q].max_val = maxv;
                        }
                    }
                }
            }
            sce->zeroes[w*16+g] = !nz;
            start += sce->ics.swb_sizes[g];
            idx++;
        }
    }
    idx--;
    mincost = paths[idx][0].cost;
    minq    = 0;
    for (i = 1; i < TRELLIS_STATES; i++) {
        if (paths[idx][i].cost < mincost) {
            mincost = paths[idx][i].cost;
            minq = i;
        }
    }
    while (idx) {
        sce->sf_idx[bandaddr[idx]] = minq;
        minq = paths[idx][minq].prev;
        idx--;
    }
    //set the same quantizers inside window groups
    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
        for (g = 0;  g < sce->ics.num_swb; g++)
            for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
                sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
}