aacdec.c

 * @param   pulse           pointer to pulse data struct
 * @param   band_type       array of the used band type
 *
 * @return  Returns error status. 0 - OK, !0 - error
 */
static int decode_spectrum_and_dequant(AACContext *ac, float coef[1024],
                                       GetBitContext *gb, const float sf[120],
                                       int pulse_present, const Pulse *pulse,
                                       const IndividualChannelStream *ics,
                                       enum BandType band_type[120])
{
    int i, k, g, idx = 0;
    const int c = 1024 / ics->num_windows;
    const uint16_t *offsets = ics->swb_offset;
    float *coef_base = coef;

    for (g = 0; g < ics->num_windows; g++)
        memset(coef + g * 128 + offsets[ics->max_sfb], 0, sizeof(float) * (c - offsets[ics->max_sfb]));

    for (g = 0; g < ics->num_window_groups; g++) {
        unsigned g_len = ics->group_len[g];

        for (i = 0; i < ics->max_sfb; i++, idx++) {
            const unsigned cbt_m1 = band_type[idx] - 1;
            float *cfo = coef + offsets[i];
            int off_len = offsets[i + 1] - offsets[i];
            int group;

            if (cbt_m1 >= INTENSITY_BT2 - 1) {
                for (group = 0; group < g_len; group++, cfo+=128) {
                    memset(cfo, 0, off_len * sizeof(float));
                }
            } else if (cbt_m1 == NOISE_BT - 1) {
                for (group = 0; group < g_len; group++, cfo+=128) {
                    float scale;
                    float band_energy;

                    for (k = 0; k < off_len; k++) {
                        ac->random_state  = lcg_random(ac->random_state);
                        cfo[k] = ac->random_state;
                    }

                    band_energy = ac->dsp.scalarproduct_float(cfo, cfo, off_len);
                    scale = sf[idx] / sqrtf(band_energy);
                    ac->dsp.vector_fmul_scalar(cfo, cfo, scale, off_len);
                }
            } else {
                const float *vq = ff_aac_codebook_vector_vals[cbt_m1];
                const uint16_t *cb_vector_idx = ff_aac_codebook_vector_idx[cbt_m1];
                VLC_TYPE (*vlc_tab)[2] = vlc_spectral[cbt_m1].table;
                OPEN_READER(re, gb);

                switch (cbt_m1 >> 1) {
                case 0:
                    for (group = 0; group < g_len; group++, cfo+=128) {
                        float *cf = cfo;
                        int len = off_len;

                        do {
                            int code;
                            unsigned cb_idx;

                            UPDATE_CACHE(re, gb);
                            GET_VLC(code, re, gb, vlc_tab, 8, 2);
                            cb_idx = cb_vector_idx[code];
                            cf = VMUL4(cf, vq, cb_idx, sf + idx);
                        } while (len -= 4);
                    }
                    break;

                case 1:
                    for (group = 0; group < g_len; group++, cfo+=128) {
                        float *cf = cfo;
                        int len = off_len;

                        do {
                            int code;
                            unsigned nnz;
                            unsigned cb_idx;
                            uint32_t bits;

                            UPDATE_CACHE(re, gb);
                            GET_VLC(code, re, gb, vlc_tab, 8, 2);
                            cb_idx = cb_vector_idx[code];
                            nnz = cb_idx >> 8 & 15;
                            bits = SHOW_UBITS(re, gb, nnz) << (32-nnz);
                            LAST_SKIP_BITS(re, gb, nnz);
                            cf = VMUL4S(cf, vq, cb_idx, bits, sf + idx);
                        } while (len -= 4);
                    }
                    break;

                case 2:
                    for (group = 0; group < g_len; group++, cfo+=128) {
                        float *cf = cfo;
                        int len = off_len;

                        do {
                            int code;
                            unsigned cb_idx;

                            UPDATE_CACHE(re, gb);
                            GET_VLC(code, re, gb, vlc_tab, 8, 2);
                            cb_idx = cb_vector_idx[code];
                            cf = VMUL2(cf, vq, cb_idx, sf + idx);
                        } while (len -= 2);
                    }
                    break;

                case 3:
                case 4:
                    for (group = 0; group < g_len; group++, cfo+=128) {
                        float *cf = cfo;
                        int len = off_len;

                        do {
                            int code;
                            unsigned nnz;
                            unsigned cb_idx;
                            unsigned sign;

                            UPDATE_CACHE(re, gb);
                            GET_VLC(code, re, gb, vlc_tab, 8, 2);
                            cb_idx = cb_vector_idx[code];
                            nnz = cb_idx >> 8 & 15;
                            sign = SHOW_UBITS(re, gb, nnz) << (cb_idx >> 12);
                            LAST_SKIP_BITS(re, gb, nnz);
                            cf = VMUL2S(cf, vq, cb_idx, sign, sf + idx);
                        } while (len -= 2);
                    }
                    break;

                default:
                    for (group = 0; group < g_len; group++, cfo+=128) {
                        float *cf = cfo;
                        uint32_t *icf = (uint32_t *) cf;
                        int len = off_len;

                        do {
                            int code;
                            unsigned nzt, nnz;
                            unsigned cb_idx;
                            uint32_t bits;
                            int j;

                            UPDATE_CACHE(re, gb);
                            GET_VLC(code, re, gb, vlc_tab, 8, 2);

                            if (!code) {
                                *icf++ = 0;
                                *icf++ = 0;
                                continue;
                            }

                            cb_idx = cb_vector_idx[code];
                            nnz = cb_idx >> 12;
                            nzt = cb_idx >> 8;
                            bits = SHOW_UBITS(re, gb, nnz) << (32-nnz);
                            LAST_SKIP_BITS(re, gb, nnz);

                            for (j = 0; j < 2; j++) {
                                if (nzt & 1<<j) {
                                    uint32_t b;
                                    int n;
                                    /* The total length of escape_sequence must be < 22 bits according
                                       to the specification (i.e. max is 111111110xxxxxxxxxxxx). */
                                    UPDATE_CACHE(re, gb);
                                    b = GET_CACHE(re, gb);
                                    b = 31 - av_log2(~b);

                                    if (b > 8) {
                                        av_log(ac->avctx, AV_LOG_ERROR, "error in spectral data, ESC overflow\n");
                                        return -1;
                                    }

                                    SKIP_BITS(re, gb, b + 1);
                                    b += 4;
                                    n = (1 << b) + SHOW_UBITS(re, gb, b);
                                    LAST_SKIP_BITS(re, gb, b);
                                    *icf++ = cbrt_tab[n] | (bits & 1U<<31);
                                    bits <<= 1;
                                } else {
                                    unsigned v = ((const uint32_t*)vq)[cb_idx & 15];
                                    *icf++ = (bits & 1U<<31) | v;
                                    bits <<= !!v;
                                }
                                cb_idx >>= 4;
                            }
                        } while (len -= 2);

                        ac->dsp.vector_fmul_scalar(cfo, cfo, sf[idx], off_len);
                    }
                }

                CLOSE_READER(re, gb);
            }
        }
        coef += g_len << 7;
    }

    if (pulse_present) {
        idx = 0;
        for (i = 0; i < pulse->num_pulse; i++) {
            float co = coef_base[ pulse->pos[i] ];
            while (offsets[idx + 1] <= pulse->pos[i])
                idx++;
            if (band_type[idx] != NOISE_BT && sf[idx]) {
                float ico = -pulse->amp[i];
                if (co) {
                    co /= sf[idx];
                    ico = co / sqrtf(sqrtf(fabsf(co))) + (co > 0 ? -ico : ico);
                }
                coef_base[ pulse->pos[i] ] = cbrtf(fabsf(ico)) * ico * sf[idx];
            }
        }
    }
    return 0;
}

static av_always_inline float flt16_round(float pf)
{
    union float754 tmp;
    tmp.f = pf;
    tmp.i = (tmp.i + 0x00008000U) & 0xFFFF0000U;
    return tmp.f;
}

static av_always_inline float flt16_even(float pf)
{
    union float754 tmp;
    tmp.f = pf;
    tmp.i = (tmp.i + 0x00007FFFU + (tmp.i & 0x00010000U >> 16)) & 0xFFFF0000U;
    return tmp.f;
}

static av_always_inline float flt16_trunc(float pf)
{
    union float754 pun;
    pun.f = pf;
    pun.i &= 0xFFFF0000U;
    return pun.f;
}

static av_always_inline void predict(PredictorState *ps, float *coef,
                                     float sf_scale, float inv_sf_scale,
                                     int output_enable)
{
    const float a     = 0.953125; // 61.0 / 64
    const float alpha = 0.90625;  // 29.0 / 32
    float e0, e1;
    float pv;
    float k1, k2;
    float   r0 = ps->r0,     r1 = ps->r1;
    float cor0 = ps->cor0, cor1 = ps->cor1;
    float var0 = ps->var0, var1 = ps->var1;

    k1 = var0 > 1 ? cor0 * flt16_even(a / var0) : 0;
    k2 = var1 > 1 ? cor1 * flt16_even(a / var1) : 0;

    pv = flt16_round(k1 * r0 + k2 * r1);
    if (output_enable)
        *coef += pv * sf_scale;

    e0 = *coef * inv_sf_scale;
    e1 = e0 - k1 * r0;

    ps->cor1 = flt16_trunc(alpha * cor1 + r1 * e1);
    ps->var1 = flt16_trunc(alpha * var1 + 0.5f * (r1 * r1 + e1 * e1));
    ps->cor0 = flt16_trunc(alpha * cor0 + r0 * e0);
    ps->var0 = flt16_trunc(alpha * var0 + 0.5f * (r0 * r0 + e0 * e0));

    ps->r1 = flt16_trunc(a * (r0 - k1 * e0));
    ps->r0 = flt16_trunc(a * e0);
}

/**
 * Apply AAC-Main style frequency domain prediction.
 */
static void apply_prediction(AACContext *ac, SingleChannelElement *sce)
{
    int sfb, k;
    float sf_scale = ac->sf_scale, inv_sf_scale = 1 / ac->sf_scale;

    if (!sce->ics.predictor_initialized) {
        reset_all_predictors(sce->predictor_state);
        sce->ics.predictor_initialized = 1;
    }

    if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
        for (sfb = 0; sfb < ff_aac_pred_sfb_max[ac->m4ac.sampling_index]; sfb++) {
            for (k = sce->ics.swb_offset[sfb]; k < sce->ics.swb_offset[sfb + 1]; k++) {
                predict(&sce->predictor_state[k], &sce->coeffs[k],
                        sf_scale, inv_sf_scale,
                        sce->ics.predictor_present && sce->ics.prediction_used[sfb]);
            }
        }
        if (sce->ics.predictor_reset_group)
            reset_predictor_group(sce->predictor_state, sce->ics.predictor_reset_group);
    } else
        reset_all_predictors(sce->predictor_state);
}

/**
 * Decode an individual_channel_stream payload; reference: table 4.44.
 *
 * @param   common_window   Channels have independent [0], or shared [1], Individual Channel Stream information.
 * @param   scale_flag      scalable [1] or non-scalable [0] AAC (Unused until scalable AAC is implemented.)
 *
 * @return  Returns error status. 0 - OK, !0 - error
 */
static int decode_ics(AACContext *ac, SingleChannelElement *sce,
                      GetBitContext *gb, int common_window, int scale_flag)
{
    Pulse pulse;
    TemporalNoiseShaping    *tns = &sce->tns;
    IndividualChannelStream *ics = &sce->ics;
    float *out = sce->coeffs;
    int global_gain, pulse_present = 0;

    /* This assignment is to silence a GCC warning about the variable being used
     * uninitialized when in fact it always is.
     */
    pulse.num_pulse = 0;

    global_gain = get_bits(gb, 8);

    if (!common_window && !scale_flag) {
        if (decode_ics_info(ac, ics, gb, 0) < 0)
            return -1;
    }

    if (decode_band_types(ac, sce->band_type, sce->band_type_run_end, gb, ics) < 0)
        return -1;
    if (decode_scalefactors(ac, sce->sf, gb, global_gain, ics, sce->band_type, sce->band_type_run_end) < 0)
        return -1;

    pulse_present = 0;
    if (!scale_flag) {
        if ((pulse_present = get_bits1(gb))) {
            if (ics->window_sequence[0] == EIGHT_SHORT_SEQUENCE) {
                av_log(ac->avctx, AV_LOG_ERROR, "Pulse tool not allowed in eight short sequence.\n");
                return -1;
            }
            if (decode_pulses(&pulse, gb, ics->swb_offset, ics->num_swb)) {
                av_log(ac->avctx, AV_LOG_ERROR, "Pulse data corrupt or invalid.\n");
                return -1;
            }
        }
        if ((tns->present = get_bits1(gb)) && decode_tns(ac, tns, gb, ics))
            return -1;
        if (get_bits1(gb)) {
            av_log_missing_feature(ac->avctx, "SSR", 1);
            return -1;
        }
    }

    if (decode_spectrum_and_dequant(ac, out, gb, sce->sf, pulse_present, &pulse, ics, sce->band_type) < 0)
        return -1;

    if (ac->m4ac.object_type == AOT_AAC_MAIN && !common_window)
        apply_prediction(ac, sce);

    return 0;
}

/**
 * Mid/Side stereo decoding; reference: 4.6.8.1.3.
 */
static void apply_mid_side_stereo(AACContext *ac, ChannelElement *cpe)
{
    const IndividualChannelStream *ics = &cpe->ch[0].ics;
    float *ch0 = cpe->ch[0].coeffs;
    float *ch1 = cpe->ch[1].coeffs;
    int g, i, group, idx = 0;
    const uint16_t *offsets = ics->swb_offset;
    for (g = 0; g < ics->num_window_groups; g++) {
        for (i = 0; i < ics->max_sfb; i++, idx++) {
            if (cpe->ms_mask[idx] &&
                    cpe->ch[0].band_type[idx] < NOISE_BT && cpe->ch[1].band_type[idx] < NOISE_BT) {
                for (group = 0; group < ics->group_len[g]; group++) {
                    ac->dsp.butterflies_float(ch0 + group * 128 + offsets[i],
                                              ch1 + group * 128 + offsets[i],
                                              offsets[i+1] - offsets[i]);
                }
            }
        }
        ch0 += ics->group_len[g] * 128;
        ch1 += ics->group_len[g] * 128;
    }
}

/**
 * intensity stereo decoding; reference: 4.6.8.2.3
 *
 * @param   ms_present  Indicates mid/side stereo presence. [0] mask is all 0s;
 *                      [1] mask is decoded from bitstream; [2] mask is all 1s;
 *                      [3] reserved for scalable AAC
 */
static void apply_intensity_stereo(AACContext *ac, ChannelElement *cpe, int ms_present)
{
    const IndividualChannelStream *ics = &cpe->ch[1].ics;
    SingleChannelElement         *sce1 = &cpe->ch[1];
    float *coef0 = cpe->ch[0].coeffs, *coef1 = cpe->ch[1].coeffs;
    const uint16_t *offsets = ics->swb_offset;
    int g, group, i, idx = 0;
    int c;
    float scale;
    for (g = 0; g < ics->num_window_groups; g++) {
        for (i = 0; i < ics->max_sfb;) {
            if (sce1->band_type[idx] == INTENSITY_BT || sce1->band_type[idx] == INTENSITY_BT2) {
                const int bt_run_end = sce1->band_type_run_end[idx];
                for (; i < bt_run_end; i++, idx++) {
                    c = -1 + 2 * (sce1->band_type[idx] - 14);
                    if (ms_present)
                        c *= 1 - 2 * cpe->ms_mask[idx];
                    scale = c * sce1->sf[idx];
                    for (group = 0; group < ics->group_len[g]; group++)
                        ac->dsp.vector_fmul_scalar(coef1 + group * 128 + offsets[i],
                                                   coef0 + group * 128 + offsets[i],
                                                   scale,
                                                   offsets[i + 1] - offsets[i]);
                }
            } else {
                int bt_run_end = sce1->band_type_run_end[idx];
                idx += bt_run_end - i;
                i    = bt_run_end;
            }
        }
        coef0 += ics->group_len[g] * 128;
        coef1 += ics->group_len[g] * 128;
    }
}

/**
 * Decode a channel_pair_element; reference: table 4.4.
 *
 * @return  Returns error status. 0 - OK, !0 - error
 */
static int decode_cpe(AACContext *ac, GetBitContext *gb, ChannelElement *cpe)
{
    int i, ret, common_window, ms_present = 0;

    common_window = get_bits1(gb);
    if (common_window) {
        if (decode_ics_info(ac, &cpe->ch[0].ics, gb, 1))
            return -1;
        i = cpe->ch[1].ics.use_kb_window[0];
        cpe->ch[1].ics = cpe->ch[0].ics;
        cpe->ch[1].ics.use_kb_window[1] = i;
        if (cpe->ch[1].ics.predictor_present && (ac->m4ac.object_type != AOT_AAC_MAIN))
            if ((cpe->ch[1].ics.ltp.present = get_bits(gb, 1)))
                decode_ltp(ac, &cpe->ch[1].ics.ltp, gb, cpe->ch[1].ics.max_sfb);
        ms_present = get_bits(gb, 2);
        if (ms_present == 3) {
            av_log(ac->avctx, AV_LOG_ERROR, "ms_present = 3 is reserved.\n");
            return -1;
        } else if (ms_present)
            decode_mid_side_stereo(cpe, gb, ms_present);
    }
    if ((ret = decode_ics(ac, &cpe->ch[0], gb, common_window, 0)))
        return ret;
    if ((ret = decode_ics(ac, &cpe->ch[1], gb, common_window, 0)))
        return ret;

    if (common_window) {
        if (ms_present)
            apply_mid_side_stereo(ac, cpe);
        if (ac->m4ac.object_type == AOT_AAC_MAIN) {
            apply_prediction(ac, &cpe->ch[0]);
            apply_prediction(ac, &cpe->ch[1]);
        }
    }

    apply_intensity_stereo(ac, cpe, ms_present);
    return 0;
}

static const float cce_scale[] = {
    1.09050773266525765921, //2^(1/8)
    1.18920711500272106672, //2^(1/4)
    M_SQRT2,
    2,
};

/**
 * Decode coupling_channel_element; reference: table 4.8.
 *
 * @return  Returns error status. 0 - OK, !0 - error
 */
static int decode_cce(AACContext *ac, GetBitContext *gb, ChannelElement *che)
{
    int num_gain = 0;
    int c, g, sfb, ret;
    int sign;
    float scale;
    SingleChannelElement *sce = &che->ch[0];
    ChannelCoupling     *coup = &che->coup;

    coup->coupling_point = 2 * get_bits1(gb);
    coup->num_coupled = get_bits(gb, 3);
    for (c = 0; c <= coup->num_coupled; c++) {
        num_gain++;
        coup->type[c] = get_bits1(gb) ? TYPE_CPE : TYPE_SCE;
        coup->id_select[c] = get_bits(gb, 4);
        if (coup->type[c] == TYPE_CPE) {
            coup->ch_select[c] = get_bits(gb, 2);
            if (coup->ch_select[c] == 3)
                num_gain++;
        } else
            coup->ch_select[c] = 2;
    }
    coup->coupling_point += get_bits1(gb) || (coup->coupling_point >> 1);

    sign  = get_bits(gb, 1);
    scale = cce_scale[get_bits(gb, 2)];

    if ((ret = decode_ics(ac, sce, gb, 0, 0)))
        return ret;

    for (c = 0; c < num_gain; c++) {
        int idx  = 0;
        int cge  = 1;
        int gain = 0;
        float gain_cache = 1.;
        if (c) {
            cge = coup->coupling_point == AFTER_IMDCT ? 1 : get_bits1(gb);
            gain = cge ? get_vlc2(gb, vlc_scalefactors.table, 7, 3) - 60: 0;
            gain_cache = powf(scale, -gain);
        }
        if (coup->coupling_point == AFTER_IMDCT) {
            coup->gain[c][0] = gain_cache;
        } else {
            for (g = 0; g < sce->ics.num_window_groups; g++) {
                for (sfb = 0; sfb < sce->ics.max_sfb; sfb++, idx++) {
                    if (sce->band_type[idx] != ZERO_BT) {
                        if (!cge) {
                            int t = get_vlc2(gb, vlc_scalefactors.table, 7, 3) - 60;
                            if (t) {
                                int s = 1;
                                t = gain += t;
                                if (sign) {
                                    s  -= 2 * (t & 0x1);
                                    t >>= 1;
                                }
                                gain_cache = powf(scale, -t) * s;
                            }
                        }
                        coup->gain[c][idx] = gain_cache;
                    }
                }
            }
        }
    }
    return 0;
}

/**
 * Parse whether channels are to be excluded from Dynamic Range Compression; reference: table 4.53.
 *
 * @return  Returns number of bytes consumed.
 */
static int decode_drc_channel_exclusions(DynamicRangeControl *che_drc,
                                         GetBitContext *gb)
{
    int i;
    int num_excl_chan = 0;

    do {
        for (i = 0; i < 7; i++)
            che_drc->exclude_mask[num_excl_chan++] = get_bits1(gb);
    } while (num_excl_chan < MAX_CHANNELS - 7 && get_bits1(gb));

    return num_excl_chan / 7;
}

/**
 * Decode dynamic range information; reference: table 4.52.
 *
 * @param   cnt length of TYPE_FIL syntactic element in bytes
 *
 * @return  Returns number of bytes consumed.
 */
static int decode_dynamic_range(DynamicRangeControl *che_drc,
                                GetBitContext *gb, int cnt)
{
    int n             = 1;
    int drc_num_bands = 1;
    int i;

    /* pce_tag_present? */
    if (get_bits1(gb)) {
        che_drc->pce_instance_tag  = get_bits(gb, 4);
        skip_bits(gb, 4); // tag_reserved_bits
        n++;
    }

    /* excluded_chns_present? */
    if (get_bits1(gb)) {
        n += decode_drc_channel_exclusions(che_drc, gb);
    }

    /* drc_bands_present? */
    if (get_bits1(gb)) {
        che_drc->band_incr            = get_bits(gb, 4);
        che_drc->interpolation_scheme = get_bits(gb, 4);
        n++;
        drc_num_bands += che_drc->band_incr;
        for (i = 0; i < drc_num_bands; i++) {
            che_drc->band_top[i] = get_bits(gb, 8);
            n++;
        }
    }

    /* prog_ref_level_present? */
    if (get_bits1(gb)) {
        che_drc->prog_ref_level = get_bits(gb, 7);
        skip_bits1(gb); // prog_ref_level_reserved_bits
        n++;
    }

    for (i = 0; i < drc_num_bands; i++) {
        che_drc->dyn_rng_sgn[i] = get_bits1(gb);
        che_drc->dyn_rng_ctl[i] = get_bits(gb, 7);
        n++;
    }

    return n;
}

/**
 * Decode extension data (incomplete); reference: table 4.51.
 *
 * @param   cnt length of TYPE_FIL syntactic element in bytes
 *
 * @return Returns number of bytes consumed
 */
static int decode_extension_payload(AACContext *ac, GetBitContext *gb, int cnt,
                                    ChannelElement *che, enum RawDataBlockType elem_type)
{
    int crc_flag = 0;
    int res = cnt;
    switch (get_bits(gb, 4)) { // extension type
    case EXT_SBR_DATA_CRC:
        crc_flag++;
    case EXT_SBR_DATA:
        if (!che) {
            av_log(ac->avctx, AV_LOG_ERROR, "SBR was found before the first channel element.\n");
            return res;
        } else if (!ac->m4ac.sbr) {
            av_log(ac->avctx, AV_LOG_ERROR, "SBR signaled to be not-present but was found in the bitstream.\n");
            skip_bits_long(gb, 8 * cnt - 4);
            return res;
        } else if (ac->m4ac.sbr == -1 && ac->output_configured == OC_LOCKED) {
            av_log(ac->avctx, AV_LOG_ERROR, "Implicit SBR was found with a first occurrence after the first frame.\n");
            skip_bits_long(gb, 8 * cnt - 4);
            return res;
        } else if (ac->m4ac.ps == -1 && ac->output_configured < OC_LOCKED && ac->avctx->channels == 1) {
            ac->m4ac.sbr = 1;
            ac->m4ac.ps = 1;
            output_configure(ac, ac->che_pos, ac->che_pos, ac->m4ac.chan_config, ac->output_configured);
        } else {
            ac->m4ac.sbr = 1;
        }
        res = ff_decode_sbr_extension(ac, &che->sbr, gb, crc_flag, cnt, elem_type);
        break;
    case EXT_DYNAMIC_RANGE:
        res = decode_dynamic_range(&ac->che_drc, gb, cnt);
        break;
    case EXT_FILL:
    case EXT_FILL_DATA:
    case EXT_DATA_ELEMENT:
    default:
        skip_bits_long(gb, 8 * cnt - 4);
        break;
    };
    return res;
}

/**
 * Decode Temporal Noise Shaping filter coefficients and apply all-pole filters; reference: 4.6.9.3.
 *
 * @param   decode  1 if tool is used normally, 0 if tool is used in LTP.
 * @param   coef    spectral coefficients
 */
static void apply_tns(float coef[1024], TemporalNoiseShaping *tns,
                      IndividualChannelStream *ics, int decode)
{
    const int mmm = FFMIN(ics->tns_max_bands, ics->max_sfb);
    int w, filt, m, i;
    int bottom, top, order, start, end, size, inc;
    float lpc[TNS_MAX_ORDER];
    float tmp[TNS_MAX_ORDER];

    for (w = 0; w < ics->num_windows; w++) {
        bottom = ics->num_swb;
        for (filt = 0; filt < tns->n_filt[w]; filt++) {
            top    = bottom;
            bottom = FFMAX(0, top - tns->length[w][filt]);
            order  = tns->order[w][filt];
            if (order == 0)
                continue;

            // tns_decode_coef
            compute_lpc_coefs(tns->coef[w][filt], order, lpc, 0, 0, 0);

            start = ics->swb_offset[FFMIN(bottom, mmm)];
            end   = ics->swb_offset[FFMIN(   top, mmm)];
            if ((size = end - start) <= 0)
                continue;
            if (tns->direction[w][filt]) {
                inc = -1;
                start = end - 1;
            } else {
                inc = 1;
            }
            start += w * 128;

            if (decode) {
                // ar filter
                for (m = 0; m < size; m++, start += inc)
                    for (i = 1; i <= FFMIN(m, order); i++)
                        coef[start] -= coef[start - i * inc] * lpc[i - 1];
            } else {
                // ma filter
                for (m = 0; m < size; m++, start += inc) {
                    tmp[0] = coef[start];
                    for (i = 1; i <= FFMIN(m, order); i++)
                        coef[start] += tmp[i] * lpc[i - 1];
                    for (i = order; i > 0; i--)
                        tmp[i] = tmp[i - 1];
                }
            }
        }
    }
}

/**
 *  Apply windowing and MDCT to obtain the spectral
 *  coefficient from the predicted sample by LTP.
 */
static void windowing_and_mdct_ltp(AACContext *ac, float *out,
                                   float *in, IndividualChannelStream *ics)
{
    const float *lwindow      = ics->use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
    const float *swindow      = ics->use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
    const float *lwindow_prev = ics->use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
    const float *swindow_prev = ics->use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;

    if (ics->window_sequence[0] != LONG_STOP_SEQUENCE) {
        ac->dsp.vector_fmul(in, in, lwindow_prev, 1024);
    } else {
        memset(in, 0, 448 * sizeof(float));
        ac->dsp.vector_fmul(in + 448, in + 448, swindow_prev, 128);
        memcpy(in + 576, in + 576, 448 * sizeof(float));
    }
    if (ics->window_sequence[0] != LONG_START_SEQUENCE) {
        ac->dsp.vector_fmul_reverse(in + 1024, in + 1024, lwindow, 1024);
    } else {
        memcpy(in + 1024, in + 1024, 448 * sizeof(float));
        ac->dsp.vector_fmul_reverse(in + 1024 + 448, in + 1024 + 448, swindow, 128);
        memset(in + 1024 + 576, 0, 448 * sizeof(float));
    }
    ac->mdct_ltp.mdct_calc(&ac->mdct_ltp, out, in);
}

/**
 * Apply the long term prediction
 */
static void apply_ltp(AACContext *ac, SingleChannelElement *sce)
{
    const LongTermPrediction *ltp = &sce->ics.ltp;
    const uint16_t *offsets = sce->ics.swb_offset;
    int i, sfb;

    if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
        float *predTime = sce->ret;
        float *predFreq = ac->buf_mdct;
        int16_t num_samples = 2048;

        if (ltp->lag < 1024)
            num_samples = ltp->lag + 1024;
        for (i = 0; i < num_samples; i++)
            predTime[i] = sce->ltp_state[i + 2048 - ltp->lag] * ltp->coef;
        memset(&predTime[i], 0, (2048 - i) * sizeof(float));

        windowing_and_mdct_ltp(ac, predFreq, predTime, &sce->ics);

        if (sce->tns.present)
            apply_tns(predFreq, &sce->tns, &sce->ics, 0);

        for (sfb = 0; sfb < FFMIN(sce->ics.max_sfb, MAX_LTP_LONG_SFB); sfb++)
            if (ltp->used[sfb])
                for (i = offsets[sfb]; i < offsets[sfb + 1]; i++)
                    sce->coeffs[i] += predFreq[i];
    }
}

/**
 * Update the LTP buffer for next frame
 */
static void update_ltp(AACContext *ac, SingleChannelElement *sce)
{
    IndividualChannelStream *ics = &sce->ics;
    float *saved     = sce->saved;
    float *saved_ltp = sce->coeffs;
    const float *lwindow = ics->use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
    const float *swindow = ics->use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
    int i;

    if (ics->window_sequence[0] == EIGHT_SHORT_SEQUENCE) {
        memcpy(saved_ltp,       saved, 512 * sizeof(float));
        memset(saved_ltp + 576, 0,     448 * sizeof(float));
        ac->dsp.vector_fmul_reverse(saved_ltp + 448, ac->buf_mdct + 960,     &swindow[64],      64);
        for (i = 0; i < 64; i++)
            saved_ltp[i + 512] = ac->buf_mdct[1023 - i] * swindow[63 - i];
    } else if (ics->window_sequence[0] == LONG_START_SEQUENCE) {
        memcpy(saved_ltp,       ac->buf_mdct + 512, 448 * sizeof(float));
        memset(saved_ltp + 576, 0,                  448 * sizeof(float));
        ac->dsp.vector_fmul_reverse(saved_ltp + 448, ac->buf_mdct + 960,     &swindow[64],      64);
        for (i = 0; i < 64; i++)
            saved_ltp[i + 512] = ac->buf_mdct[1023 - i] * swindow[63 - i];
    } else { // LONG_STOP or ONLY_LONG
        ac->dsp.vector_fmul_reverse(saved_ltp,       ac->buf_mdct + 512,     &lwindow[512],     512);
        for (i = 0; i < 512; i++)
            saved_ltp[i + 512] = ac->buf_mdct[1023 - i] * lwindow[511 - i];
    }

    memcpy(sce->ltp_state, &sce->ltp_state[1024], 1024 * sizeof(int16_t));
    ac->fmt_conv.float_to_int16(&(sce->ltp_state[1024]), sce->ret,  1024);
    ac->fmt_conv.float_to_int16(&(sce->ltp_state[2048]), saved_ltp, 1024);
}

/**
 * Conduct IMDCT and windowing.
 */
static void imdct_and_windowing(AACContext *ac, SingleChannelElement *sce)
{
    IndividualChannelStream *ics = &sce->ics;
    float *in    = sce->coeffs;
    float *out   = sce->ret;
    float *saved = sce->saved;
    const float *swindow      = ics->use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
    const float *lwindow_prev = ics->use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
    const float *swindow_prev = ics->use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
    float *buf  = ac->buf_mdct;
    float *temp = ac->temp;
    int i;

    // imdct
    if (ics->window_sequence[0] == EIGHT_SHORT_SEQUENCE) {
        for (i = 0; i < 1024; i += 128)
            ac->mdct_small.imdct_half(&ac->mdct_small, buf + i, in + i);
    } else
        ac->mdct.imdct_half(&ac->mdct, buf, in);

    /* window overlapping
     * NOTE: To simplify the overlapping code, all 'meaningless' short to long
     * and long to short transitions are considered to be short to short
     * transitions. This leaves just two cases (long to long and short to short)
     * with a little special sauce for EIGHT_SHORT_SEQUENCE.
     */
    if ((ics->window_sequence[1] == ONLY_LONG_SEQUENCE || ics->window_sequence[1] == LONG_STOP_SEQUENCE) &&
            (ics->window_sequence[0] == ONLY_LONG_SEQUENCE || ics->window_sequence[0] == LONG_START_SEQUENCE)) {
        ac->dsp.vector_fmul_window(    out,               saved,            buf,         lwindow_prev, 512);
    } else {
        memcpy(                        out,               saved,            448 * sizeof(float));

        if (ics->window_sequence[0] == EIGHT_SHORT_SEQUENCE) {
            ac->dsp.vector_fmul_window(out + 448 + 0*128, saved + 448,      buf + 0*128, swindow_prev, 64);
            ac->dsp.vector_fmul_window(out + 448 + 1*128, buf + 0*128 + 64, buf + 1*128, swindow,      64);
            ac->dsp.vector_fmul_window(out + 448 + 2*128, buf + 1*128 + 64, buf + 2*128, swindow,      64);
            ac->dsp.vector_fmul_window(out + 448 + 3*128, buf + 2*128 + 64, buf + 3*128, swindow,      64);
            ac->dsp.vector_fmul_window(temp,              buf + 3*128 + 64, buf + 4*128, swindow,      64);
            memcpy(                    out + 448 + 4*128, temp, 64 * sizeof(float));
        } else {
            ac->dsp.vector_fmul_window(out + 448,         saved + 448,      buf,         swindow_prev, 64);
            memcpy(                    out + 576,         buf + 64,         448 * sizeof(float));
        }
    }

    // buffer update
    if (ics->window_sequence[0] == EIGHT_SHORT_SEQUENCE) {
        memcpy(                    saved,       temp + 64,         64 * sizeof(float));
        ac->dsp.vector_fmul_window(saved + 64,  buf + 4*128 + 64, buf + 5*128, swindow, 64);
        ac->dsp.vector_fmul_window(saved + 192, buf + 5*128 + 64, buf + 6*128, swindow, 64);
        ac->dsp.vector_fmul_window(saved + 320, buf + 6*128 + 64, buf + 7*128, swindow, 64);
        memcpy(                    saved + 448, buf + 7*128 + 64,  64 * sizeof(float));
    } else if (ics->window_sequence[0] == LONG_START_SEQUENCE) {
        memcpy(                    saved,       buf + 512,        448 * sizeof(float));
        memcpy(                    saved + 448, buf + 7*128 + 64,  64 * sizeof(float));
    } else { // LONG_STOP or ONLY_LONG
        memcpy(                    saved,       buf + 512,        512 * sizeof(float));
    }
}

/**
 * Apply dependent channel coupling (applied before IMDCT).
 *
 * @param   index   index into coupling gain array
 */
static void apply_dependent_coupling(AACContext *ac,
                                     SingleChannelElement *target,
                                     ChannelElement *cce, int index)
{
    IndividualChannelStream *ics = &cce->ch[0].ics;
    const uint16_t *offsets = ics->swb_offset;
    float *dest = target->coeffs;
    const float *src = cce->ch[0].coeffs;
    int g, i, group, k, idx = 0;
    if (ac->m4ac.object_type == AOT_AAC_LTP) {
        av_log(ac->avctx, AV_LOG_ERROR,
               "Dependent coupling is not supported together with LTP\n");
        return;
    }
    for (g = 0; g < ics->num_window_groups; g++) {
        for (i = 0; i < ics->max_sfb; i++, idx++) {
            if (cce->ch[0].band_type[idx] != ZERO_BT) {
                const float gain = cce->coup.gain[index][idx];
                for (group = 0; group < ics->group_len[g]; group++) {
                    for (k = offsets[i]; k < offsets[i + 1]; k++) {
                        // XXX dsputil-ize
                        dest[group * 128 + k] += gain * src[group * 128 + k];
                    }
                }
            }
        }
        dest += ics->group_len[g] * 128;
        src  += ics->group_len[g] * 128;
    }
}

/**
 * Apply independent channel coupling (applied after IMDCT).
 *
 * @param   index   index into coupling gain array
 */
static void apply_independent_coupling(AACContext *ac,
                                       SingleChannelElement *target,
                                       ChannelElement *cce, int index)
{
    int i;
    const float gain = cce->coup.gain[index][0];
    const float *src = cce->ch[0].ret;
    float *dest = target->ret;
    const int len = 1024 << (ac->m4ac.sbr == 1);

    for (i = 0; i < len; i++)
        dest[i] += gain * src[i];
}

/**
 * channel coupling transformation interface
 *
 * @param   apply_coupling_method   pointer to (in)dependent coupling function
 */
static void apply_channel_coupling(AACContext *ac, ChannelElement *cc,
                                   enum RawDataBlockType type, int elem_id,
                                   enum CouplingPoint coupling_point,
                                   void (*apply_coupling_method)(AACContext *ac, SingleChannelElement *target, ChannelElement *cce, int index))
{
    int i, c;

    for (i = 0; i < MAX_ELEM_ID; i++) {
        ChannelElement *cce = ac->che[TYPE_CCE][i];
        int index = 0;

        if (cce && cce->coup.coupling_point == coupling_point) {
            ChannelCoupling *coup = &cce->coup;

            for (c = 0; c <= coup->num_coupled; c++) {
                if (coup->type[c] == type && coup->id_select[c] == elem_id) {
                    if (coup->ch_select[c] != 1) {
                        apply_coupling_method(ac, &cc->ch[0], cce, index);
                        if (coup->ch_select[c] != 0)
                            index++;
                    }
                    if (coup->ch_select[c] != 2)
                        apply_coupling_method(ac, &cc->ch[1], cce, index++);
                } else
                    index += 1 + (coup->ch_select[c] == 3);
            }
        }
    }
}

/**
 * Convert spectral data to float samples, applying all supported tools as appropriate.
 */
static void spectral_to_sample(AACContext *ac)
{
    int i, type;
    for (type = 3; type >= 0; type--) {
        for (i = 0; i < MAX_ELEM_ID; i++) {
            ChannelElement *che = ac->che[type][i];
            if (che) {
                if (type <= TYPE_CPE)
                    apply_channel_coupling(ac, che, type, i, BEFORE_TNS, apply_dependent_coupling);
                if (ac->m4ac.object_type == AOT_AAC_LTP) {
                    if (che->ch[0].ics.predictor_present) {
                        if (che->ch[0].ics.ltp.present)