FFmpeg/libavcodec/g729dec.c

/*
 * G.729, G729 Annex D decoders
 * Copyright (c) 2008 Vladimir Voroshilov
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

#include <inttypes.h>
#include <string.h>

#include "avcodec.h"
#include "libavutil/avutil.h"
#include "get_bits.h"
#include "audiodsp.h"
#include "internal.h"


#include "g729.h"
#include "lsp.h"
#include "celp_math.h"
#include "celp_filters.h"
#include "acelp_filters.h"
#include "acelp_pitch_delay.h"
#include "acelp_vectors.h"
#include "g729data.h"
#include "g729postfilter.h"

/**
 * minimum quantized LSF value (3.2.4)
 * 0.005 in Q13
 */
#define LSFQ_MIN                   40

/**
 * maximum quantized LSF value (3.2.4)
 * 3.135 in Q13
 */
#define LSFQ_MAX                   25681

/**
 * minimum LSF distance (3.2.4)
 * 0.0391 in Q13
 */
#define LSFQ_DIFF_MIN              321

/// interpolation filter length
#define INTERPOL_LEN              11

/**
 * minimum gain pitch value (3.8, Equation 47)
 * 0.2 in (1.14)
 */
#define SHARP_MIN                  3277

/**
 * maximum gain pitch value (3.8, Equation 47)
 * (EE) This does not comply with the specification.
 * Specification says about 0.8, which should be
 * 13107 in (1.14), but reference C code uses
 * 13017 (equals to 0.7945) instead of it.
 */
#define SHARP_MAX                  13017

/**
 * MR_ENERGY (mean removed energy) = mean_energy + 10 * log10(2^26  * subframe_size) in (7.13)
 */
#define MR_ENERGY 1018156

#define DECISION_NOISE        0
#define DECISION_INTERMEDIATE 1
#define DECISION_VOICE        2

typedef enum {
    FORMAT_G729_8K = 0,
    FORMAT_G729D_6K4,
    FORMAT_COUNT,
} G729Formats;

typedef struct {
    uint8_t ac_index_bits[2];   ///< adaptive codebook index for second subframe (size in bits)
    uint8_t parity_bit;         ///< parity bit for pitch delay
    uint8_t gc_1st_index_bits;  ///< gain codebook (first stage) index (size in bits)
    uint8_t gc_2nd_index_bits;  ///< gain codebook (second stage) index (size in bits)
    uint8_t fc_signs_bits;      ///< number of pulses in fixed-codebook vector
    uint8_t fc_indexes_bits;    ///< size (in bits) of fixed-codebook index entry
    uint8_t block_size;
} G729FormatDescription;

typedef struct {
    /// past excitation signal buffer
    int16_t exc_base[2*SUBFRAME_SIZE+PITCH_DELAY_MAX+INTERPOL_LEN];

    int16_t* exc;               ///< start of past excitation data in buffer
    int pitch_delay_int_prev;   ///< integer part of previous subframe's pitch delay (4.1.3)

    /// (2.13) LSP quantizer outputs
    int16_t  past_quantizer_output_buf[MA_NP + 1][10];
    int16_t* past_quantizer_outputs[MA_NP + 1];

    int16_t lsfq[10];           ///< (2.13) quantized LSF coefficients from previous frame
    int16_t lsp_buf[2][10];     ///< (0.15) LSP coefficients (previous and current frames) (3.2.5)
    int16_t *lsp[2];            ///< pointers to lsp_buf

    int16_t quant_energy[4];    ///< (5.10) past quantized energy

    /// previous speech data for LP synthesis filter
    int16_t syn_filter_data[10];


    /// residual signal buffer (used in long-term postfilter)
    int16_t residual[SUBFRAME_SIZE + RES_PREV_DATA_SIZE];

    /// previous speech data for residual calculation filter
    int16_t res_filter_data[SUBFRAME_SIZE+10];

    /// previous speech data for short-term postfilter
    int16_t pos_filter_data[SUBFRAME_SIZE+10];

    /// (1.14) pitch gain of current and five previous subframes
    int16_t past_gain_pitch[6];

    /// (14.1) gain code from current and previous subframe
    int16_t past_gain_code[2];

    /// voice decision on previous subframe (0-noise, 1-intermediate, 2-voice), G.729D
    int16_t voice_decision;

    int16_t onset;              ///< detected onset level (0-2)
    int16_t was_periodic;       ///< whether previous frame was declared as periodic or not (4.4)
    int16_t ht_prev_data;       ///< previous data for 4.2.3, equation 86
    int gain_coeff;             ///< (1.14) gain coefficient (4.2.4)
    uint16_t rand_value;        ///< random number generator value (4.4.4)
    int ma_predictor_prev;      ///< switched MA predictor of LSP quantizer from last good frame

    /// (14.14) high-pass filter data (past input)
    int hpf_f[2];

    /// high-pass filter data (past output)
    int16_t hpf_z[2];
}  G729ChannelContext;

typedef struct {
    AudioDSPContext adsp;

    G729ChannelContext *channel_context;
} G729Context;

static const G729FormatDescription format_g729_8k = {
    .ac_index_bits     = {8,5},
    .parity_bit        = 1,
    .gc_1st_index_bits = GC_1ST_IDX_BITS_8K,
    .gc_2nd_index_bits = GC_2ND_IDX_BITS_8K,
    .fc_signs_bits     = 4,
    .fc_indexes_bits   = 13,
    .block_size        = G729_8K_BLOCK_SIZE,
};

static const G729FormatDescription format_g729d_6k4 = {
    .ac_index_bits     = {8,4},
    .parity_bit        = 0,
    .gc_1st_index_bits = GC_1ST_IDX_BITS_6K4,
    .gc_2nd_index_bits = GC_2ND_IDX_BITS_6K4,
    .fc_signs_bits     = 2,
    .fc_indexes_bits   = 9,
    .block_size        = G729D_6K4_BLOCK_SIZE,
};

/**
 * @brief pseudo random number generator
 */
static inline uint16_t g729_prng(uint16_t value)
{
    return 31821 * value + 13849;
}

/**
 * Decodes LSF (Line Spectral Frequencies) from L0-L3 (3.2.4).
 * @param[out] lsfq (2.13) quantized LSF coefficients
 * @param[in,out] past_quantizer_outputs (2.13) quantizer outputs from previous frames
 * @param ma_predictor switched MA predictor of LSP quantizer
 * @param vq_1st first stage vector of quantizer
 * @param vq_2nd_low second stage lower vector of LSP quantizer
 * @param vq_2nd_high second stage higher vector of LSP quantizer
 */
static void lsf_decode(int16_t* lsfq, int16_t* past_quantizer_outputs[MA_NP + 1],
                       int16_t ma_predictor,
                       int16_t vq_1st, int16_t vq_2nd_low, int16_t vq_2nd_high)
{
    int i,j;
    static const uint8_t min_distance[2]={10, 5}; //(2.13)
    int16_t* quantizer_output = past_quantizer_outputs[MA_NP];

    for (i = 0; i < 5; i++) {
        quantizer_output[i]     = cb_lsp_1st[vq_1st][i    ] + cb_lsp_2nd[vq_2nd_low ][i    ];
        quantizer_output[i + 5] = cb_lsp_1st[vq_1st][i + 5] + cb_lsp_2nd[vq_2nd_high][i + 5];
    }

    for (j = 0; j < 2; j++) {
        for (i = 1; i < 10; i++) {
            int diff = (quantizer_output[i - 1] - quantizer_output[i] + min_distance[j]) >> 1;
            if (diff > 0) {
                quantizer_output[i - 1] -= diff;
                quantizer_output[i    ] += diff;
            }
        }
    }

    for (i = 0; i < 10; i++) {
        int sum = quantizer_output[i] * cb_ma_predictor_sum[ma_predictor][i];
        for (j = 0; j < MA_NP; j++)
            sum += past_quantizer_outputs[j][i] * cb_ma_predictor[ma_predictor][j][i];

        lsfq[i] = sum >> 15;
    }

    ff_acelp_reorder_lsf(lsfq, LSFQ_DIFF_MIN, LSFQ_MIN, LSFQ_MAX, 10);
}

/**
 * Restores past LSP quantizer output using LSF from previous frame
 * @param[in,out] lsfq (2.13) quantized LSF coefficients
 * @param[in,out] past_quantizer_outputs (2.13) quantizer outputs from previous frames
 * @param ma_predictor_prev MA predictor from previous frame
 * @param lsfq_prev (2.13) quantized LSF coefficients from previous frame
 */
static void lsf_restore_from_previous(int16_t* lsfq,
                                      int16_t* past_quantizer_outputs[MA_NP + 1],
                                      int ma_predictor_prev)
{
    int16_t* quantizer_output = past_quantizer_outputs[MA_NP];
    int i,k;

    for (i = 0; i < 10; i++) {
        int tmp = lsfq[i] << 15;

        for (k = 0; k < MA_NP; k++)
            tmp -= past_quantizer_outputs[k][i] * cb_ma_predictor[ma_predictor_prev][k][i];

        quantizer_output[i] = ((tmp >> 15) * cb_ma_predictor_sum_inv[ma_predictor_prev][i]) >> 12;
    }
}

/**
 * Constructs new excitation signal and applies phase filter to it
 * @param[out] out constructed speech signal
 * @param in original excitation signal
 * @param fc_cur (2.13) original fixed-codebook vector
 * @param gain_code (14.1) gain code
 * @param subframe_size length of the subframe
 */
static void g729d_get_new_exc(
        int16_t* out,
        const int16_t* in,
        const int16_t* fc_cur,
        int dstate,
        int gain_code,
        int subframe_size)
{
    int i;
    int16_t fc_new[SUBFRAME_SIZE];

    ff_celp_convolve_circ(fc_new, fc_cur, phase_filter[dstate], subframe_size);

    for (i = 0; i < subframe_size; i++) {
        out[i]  = in[i];
        out[i] -= (gain_code * fc_cur[i] + 0x2000) >> 14;
        out[i] += (gain_code * fc_new[i] + 0x2000) >> 14;
    }
}

/**
 * Makes decision about onset in current subframe
 * @param past_onset decision result of previous subframe
 * @param past_gain_code gain code of current and previous subframe
 *
 * @return onset decision result for current subframe
 */
static int g729d_onset_decision(int past_onset, const int16_t* past_gain_code)
{
    if ((past_gain_code[0] >> 1) > past_gain_code[1])
        return 2;

    return FFMAX(past_onset-1, 0);
}

/**
 * Makes decision about voice presence in current subframe
 * @param onset onset level
 * @param prev_voice_decision voice decision result from previous subframe
 * @param past_gain_pitch pitch gain of current and previous subframes
 *
 * @return voice decision result for current subframe
 */
static int16_t g729d_voice_decision(int onset, int prev_voice_decision, const int16_t* past_gain_pitch)
{
    int i, low_gain_pitch_cnt, voice_decision;

    if (past_gain_pitch[0] >= 14745) {       // 0.9
        voice_decision = DECISION_VOICE;
    } else if (past_gain_pitch[0] <= 9830) { // 0.6
        voice_decision = DECISION_NOISE;
    } else {
        voice_decision = DECISION_INTERMEDIATE;
    }

    for (i = 0, low_gain_pitch_cnt = 0; i < 6; i++)
        if (past_gain_pitch[i] < 9830)
            low_gain_pitch_cnt++;

    if (low_gain_pitch_cnt > 2 && !onset)
        voice_decision = DECISION_NOISE;

    if (!onset && voice_decision > prev_voice_decision + 1)
        voice_decision--;

    if (onset && voice_decision < DECISION_VOICE)
        voice_decision++;

    return voice_decision;
}

static int32_t scalarproduct_int16_c(const int16_t * v1, const int16_t * v2, int order)
{
    int64_t res = 0;

    while (order--)
        res += *v1++ * *v2++;

    if      (res > INT32_MAX) return INT32_MAX;
    else if (res < INT32_MIN) return INT32_MIN;

    return res;
}

static av_cold int decoder_init(AVCodecContext * avctx)
{
    G729Context *s = avctx->priv_data;
    G729ChannelContext *ctx;
    int c,i,k;

    if (avctx->channels < 1 || avctx->channels > 2) {
        av_log(avctx, AV_LOG_ERROR, "Only mono and stereo are supported (requested channels: %d).\n", avctx->channels);
        return AVERROR(EINVAL);
    }
    avctx->sample_fmt = AV_SAMPLE_FMT_S16P;

    /* Both 8kbit/s and 6.4kbit/s modes uses two subframes per frame. */
    avctx->frame_size = SUBFRAME_SIZE << 1;

    ctx =
    s->channel_context = av_mallocz(sizeof(G729ChannelContext) * avctx->channels);
    if (!ctx)
        return AVERROR(ENOMEM);

    for (c = 0; c < avctx->channels; c++) {
        ctx->gain_coeff = 16384; // 1.0 in (1.14)

        for (k = 0; k < MA_NP + 1; k++) {
            ctx->past_quantizer_outputs[k] = ctx->past_quantizer_output_buf[k];
            for (i = 1; i < 11; i++)
                ctx->past_quantizer_outputs[k][i - 1] = (18717 * i) >> 3;
        }

        ctx->lsp[0] = ctx->lsp_buf[0];
        ctx->lsp[1] = ctx->lsp_buf[1];
        memcpy(ctx->lsp[0], lsp_init, 10 * sizeof(int16_t));

        ctx->exc = &ctx->exc_base[PITCH_DELAY_MAX+INTERPOL_LEN];

        ctx->pitch_delay_int_prev = PITCH_DELAY_MIN;

        /* random seed initialization */
        ctx->rand_value = 21845;

        /* quantized prediction error */
        for (i = 0; i < 4; i++)
            ctx->quant_energy[i] = -14336; // -14 in (5.10)

        ctx++;
    }

    ff_audiodsp_init(&s->adsp);
    s->adsp.scalarproduct_int16 = scalarproduct_int16_c;

    return 0;
}

static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame_ptr,
                        AVPacket *avpkt)
{
    const uint8_t *buf = avpkt->data;
    int buf_size       = avpkt->size;
    int16_t *out_frame;
    GetBitContext gb;
    const G729FormatDescription *format;
    int c, i;
    int16_t *tmp;
    G729Formats packet_type;
    G729Context *s = avctx->priv_data;
    G729ChannelContext *ctx = s->channel_context;
    int16_t lp[2][11];           // (3.12)
    uint8_t ma_predictor;     ///< switched MA predictor of LSP quantizer
    uint8_t quantizer_1st;    ///< first stage vector of quantizer
    uint8_t quantizer_2nd_lo; ///< second stage lower vector of quantizer (size in bits)
    uint8_t quantizer_2nd_hi; ///< second stage higher vector of quantizer (size in bits)

    int pitch_delay_int[2];      // pitch delay, integer part
    int pitch_delay_3x;          // pitch delay, multiplied by 3
    int16_t fc[SUBFRAME_SIZE];   // fixed-codebook vector
    int16_t synth[SUBFRAME_SIZE+10]; // fixed-codebook vector
    int j, ret;
    int gain_before, gain_after;
    AVFrame *frame = data;

    frame->nb_samples = SUBFRAME_SIZE<<1;
    if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
        return ret;

    if (buf_size && buf_size % ((G729_8K_BLOCK_SIZE + (avctx->codec_id == AV_CODEC_ID_ACELP_KELVIN)) * avctx->channels) == 0) {
        packet_type = FORMAT_G729_8K;
        format = &format_g729_8k;
        //Reset voice decision
        ctx->onset = 0;
        ctx->voice_decision = DECISION_VOICE;
        av_log(avctx, AV_LOG_DEBUG, "Packet type: %s\n", "G.729 @ 8kbit/s");
    } else if (buf_size == G729D_6K4_BLOCK_SIZE * avctx->channels && avctx->codec_id != AV_CODEC_ID_ACELP_KELVIN) {
        packet_type = FORMAT_G729D_6K4;
        format = &format_g729d_6k4;
        av_log(avctx, AV_LOG_DEBUG, "Packet type: %s\n", "G.729D @ 6.4kbit/s");
    } else {
        av_log(avctx, AV_LOG_ERROR, "Packet size %d is unknown.\n", buf_size);
        return AVERROR_INVALIDDATA;
    }

    for (c = 0; c < avctx->channels; c++) {
        int frame_erasure = 0; ///< frame erasure detected during decoding
        int bad_pitch = 0;     ///< parity check failed
        int is_periodic = 0;   ///< whether one of the subframes is declared as periodic or not
        out_frame = (int16_t*)frame->data[c];
        if (avctx->codec_id == AV_CODEC_ID_ACELP_KELVIN) {
            if (*buf != ((avctx->channels - 1 - c) * 0x80 | 2))
                avpriv_request_sample(avctx, "First byte value %x for channel %d", *buf, c);
            buf++;
        }

        for (i = 0; i < format->block_size; i++)
            frame_erasure |= buf[i];
        frame_erasure = !frame_erasure;

        init_get_bits8(&gb, buf, format->block_size);

        ma_predictor     = get_bits(&gb, 1);
        quantizer_1st    = get_bits(&gb, VQ_1ST_BITS);
        quantizer_2nd_lo = get_bits(&gb, VQ_2ND_BITS);
        quantizer_2nd_hi = get_bits(&gb, VQ_2ND_BITS);

        if (frame_erasure) {
            lsf_restore_from_previous(ctx->lsfq, ctx->past_quantizer_outputs,
                                      ctx->ma_predictor_prev);
        } else {
            lsf_decode(ctx->lsfq, ctx->past_quantizer_outputs,
                       ma_predictor,
                       quantizer_1st, quantizer_2nd_lo, quantizer_2nd_hi);
            ctx->ma_predictor_prev = ma_predictor;
        }

        tmp = ctx->past_quantizer_outputs[MA_NP];
        memmove(ctx->past_quantizer_outputs + 1, ctx->past_quantizer_outputs,
                MA_NP * sizeof(int16_t*));
        ctx->past_quantizer_outputs[0] = tmp;

        ff_acelp_lsf2lsp(ctx->lsp[1], ctx->lsfq, 10);

        ff_acelp_lp_decode(&lp[0][0], &lp[1][0], ctx->lsp[1], ctx->lsp[0], 10);

        FFSWAP(int16_t*, ctx->lsp[1], ctx->lsp[0]);

        for (i = 0; i < 2; i++) {
            int gain_corr_factor;

            uint8_t ac_index;      ///< adaptive codebook index
            uint8_t pulses_signs;  ///< fixed-codebook vector pulse signs
            int fc_indexes;        ///< fixed-codebook indexes
            uint8_t gc_1st_index;  ///< gain codebook (first stage) index
            uint8_t gc_2nd_index;  ///< gain codebook (second stage) index

            ac_index      = get_bits(&gb, format->ac_index_bits[i]);
            if (!i && format->parity_bit)
                bad_pitch = av_parity(ac_index >> 2) == get_bits1(&gb);
            fc_indexes    = get_bits(&gb, format->fc_indexes_bits);
            pulses_signs  = get_bits(&gb, format->fc_signs_bits);
            gc_1st_index  = get_bits(&gb, format->gc_1st_index_bits);
            gc_2nd_index  = get_bits(&gb, format->gc_2nd_index_bits);

            if (frame_erasure) {
                pitch_delay_3x = 3 * ctx->pitch_delay_int_prev;
            } else if (!i) {
                if (bad_pitch) {
                    pitch_delay_3x = 3 * ctx->pitch_delay_int_prev;
                } else {
                    pitch_delay_3x = ff_acelp_decode_8bit_to_1st_delay3(ac_index);
                }
            } else {
                int pitch_delay_min = av_clip(ctx->pitch_delay_int_prev - 5,
                                              PITCH_DELAY_MIN, PITCH_DELAY_MAX - 9);

                if (packet_type == FORMAT_G729D_6K4) {
                    pitch_delay_3x = ff_acelp_decode_4bit_to_2nd_delay3(ac_index, pitch_delay_min);
                } else {
                    pitch_delay_3x = ff_acelp_decode_5_6_bit_to_2nd_delay3(ac_index, pitch_delay_min);
                }
            }

            /* Round pitch delay to nearest (used everywhere except ff_acelp_interpolate). */
            pitch_delay_int[i]  = (pitch_delay_3x + 1) / 3;
            if (pitch_delay_int[i] > PITCH_DELAY_MAX) {
                av_log(avctx, AV_LOG_WARNING, "pitch_delay_int %d is too large\n", pitch_delay_int[i]);
                pitch_delay_int[i] = PITCH_DELAY_MAX;
            }

            if (frame_erasure) {
                ctx->rand_value = g729_prng(ctx->rand_value);
                fc_indexes   = av_mod_uintp2(ctx->rand_value, format->fc_indexes_bits);

                ctx->rand_value = g729_prng(ctx->rand_value);
                pulses_signs = ctx->rand_value;
            }


            memset(fc, 0, sizeof(int16_t) * SUBFRAME_SIZE);
            switch (packet_type) {
                case FORMAT_G729_8K:
                    ff_acelp_fc_pulse_per_track(fc, ff_fc_4pulses_8bits_tracks_13,
                                                ff_fc_4pulses_8bits_track_4,
                                                fc_indexes, pulses_signs, 3, 3);
                    break;
                case FORMAT_G729D_6K4:
                    ff_acelp_fc_pulse_per_track(fc, ff_fc_2pulses_9bits_track1_gray,
                                                ff_fc_2pulses_9bits_track2_gray,
                                                fc_indexes, pulses_signs, 1, 4);
                    break;
            }

            /*
              This filter enhances harmonic components of the fixed-codebook vector to
              improve the quality of the reconstructed speech.

                         / fc_v[i],                                    i < pitch_delay
              fc_v[i] = <
                         \ fc_v[i] + gain_pitch * fc_v[i-pitch_delay], i >= pitch_delay
            */
            if (SUBFRAME_SIZE > pitch_delay_int[i])
                ff_acelp_weighted_vector_sum(fc + pitch_delay_int[i],
                                             fc + pitch_delay_int[i],
                                             fc, 1 << 14,
                                             av_clip(ctx->past_gain_pitch[0], SHARP_MIN, SHARP_MAX),
                                             0, 14,
                                             SUBFRAME_SIZE - pitch_delay_int[i]);

            memmove(ctx->past_gain_pitch+1, ctx->past_gain_pitch, 5 * sizeof(int16_t));
            ctx->past_gain_code[1] = ctx->past_gain_code[0];

            if (frame_erasure) {
                ctx->past_gain_pitch[0] = (29491 * ctx->past_gain_pitch[0]) >> 15; // 0.90 (0.15)
                ctx->past_gain_code[0]  = ( 2007 * ctx->past_gain_code[0] ) >> 11; // 0.98 (0.11)

                gain_corr_factor = 0;
            } else {
                if (packet_type == FORMAT_G729D_6K4) {
                    ctx->past_gain_pitch[0]  = cb_gain_1st_6k4[gc_1st_index][0] +
                                               cb_gain_2nd_6k4[gc_2nd_index][0];
                    gain_corr_factor = cb_gain_1st_6k4[gc_1st_index][1] +
                                       cb_gain_2nd_6k4[gc_2nd_index][1];

                    /* Without check below overflow can occur in ff_acelp_update_past_gain.
                       It is not issue for G.729, because gain_corr_factor in it's case is always
                       greater than 1024, while in G.729D it can be even zero. */
                    gain_corr_factor = FFMAX(gain_corr_factor, 1024);
    #ifndef G729_BITEXACT
                    gain_corr_factor >>= 1;
    #endif
                } else {
                    ctx->past_gain_pitch[0]  = cb_gain_1st_8k[gc_1st_index][0] +
                                               cb_gain_2nd_8k[gc_2nd_index][0];
                    gain_corr_factor = cb_gain_1st_8k[gc_1st_index][1] +
                                       cb_gain_2nd_8k[gc_2nd_index][1];
                }

                /* Decode the fixed-codebook gain. */
                ctx->past_gain_code[0] = ff_acelp_decode_gain_code(&s->adsp, gain_corr_factor,
                                                                   fc, MR_ENERGY,
                                                                   ctx->quant_energy,
                                                                   ma_prediction_coeff,
                                                                   SUBFRAME_SIZE, 4);
    #ifdef G729_BITEXACT
                /*
                  This correction required to get bit-exact result with
                  reference code, because gain_corr_factor in G.729D is
                  two times larger than in original G.729.

                  If bit-exact result is not issue then gain_corr_factor
                  can be simpler divided by 2 before call to g729_get_gain_code
                  instead of using correction below.
                */
                if (packet_type == FORMAT_G729D_6K4) {
                    gain_corr_factor >>= 1;
                    ctx->past_gain_code[0] >>= 1;
                }
    #endif
            }
            ff_acelp_update_past_gain(ctx->quant_energy, gain_corr_factor, 2, frame_erasure);

            /* Routine requires rounding to lowest. */
            ff_acelp_interpolate(ctx->exc + i * SUBFRAME_SIZE,
                                 ctx->exc + i * SUBFRAME_SIZE - pitch_delay_3x / 3,
                                 ff_acelp_interp_filter, 6,
                                 (pitch_delay_3x % 3) << 1,
                                 10, SUBFRAME_SIZE);

            ff_acelp_weighted_vector_sum(ctx->exc + i * SUBFRAME_SIZE,
                                         ctx->exc + i * SUBFRAME_SIZE, fc,
                                         (!ctx->was_periodic && frame_erasure) ? 0 : ctx->past_gain_pitch[0],
                                         ( ctx->was_periodic && frame_erasure) ? 0 : ctx->past_gain_code[0],
                                         1 << 13, 14, SUBFRAME_SIZE);

            memcpy(synth, ctx->syn_filter_data, 10 * sizeof(int16_t));

            if (ff_celp_lp_synthesis_filter(
                synth+10,
                &lp[i][1],
                ctx->exc  + i * SUBFRAME_SIZE,
                SUBFRAME_SIZE,
                10,
                1,
                0,
                0x800))
                /* Overflow occurred, downscale excitation signal... */
                for (j = 0; j < 2 * SUBFRAME_SIZE + PITCH_DELAY_MAX + INTERPOL_LEN; j++)
                    ctx->exc_base[j] >>= 2;

            /* ... and make synthesis again. */
            if (packet_type == FORMAT_G729D_6K4) {
                int16_t exc_new[SUBFRAME_SIZE];

                ctx->onset = g729d_onset_decision(ctx->onset, ctx->past_gain_code);
                ctx->voice_decision = g729d_voice_decision(ctx->onset, ctx->voice_decision, ctx->past_gain_pitch);

                g729d_get_new_exc(exc_new, ctx->exc  + i * SUBFRAME_SIZE, fc, ctx->voice_decision, ctx->past_gain_code[0], SUBFRAME_SIZE);

                ff_celp_lp_synthesis_filter(
                        synth+10,
                        &lp[i][1],
                        exc_new,
                        SUBFRAME_SIZE,
                        10,
                        0,
                        0,
                        0x800);
            } else {
                ff_celp_lp_synthesis_filter(
                        synth+10,
                        &lp[i][1],
                        ctx->exc  + i * SUBFRAME_SIZE,
                        SUBFRAME_SIZE,
                        10,
                        0,
                        0,
                        0x800);
            }
            /* Save data (without postfilter) for use in next subframe. */
            memcpy(ctx->syn_filter_data, synth+SUBFRAME_SIZE, 10 * sizeof(int16_t));

            /* Calculate gain of unfiltered signal for use in AGC. */
            gain_before = 0;
            for (j = 0; j < SUBFRAME_SIZE; j++)
                gain_before += FFABS(synth[j+10]);

            /* Call postfilter and also update voicing decision for use in next frame. */
            ff_g729_postfilter(
                    &s->adsp,
                    &ctx->ht_prev_data,
                    &is_periodic,
                    &lp[i][0],
                    pitch_delay_int[0],
                    ctx->residual,
                    ctx->res_filter_data,
                    ctx->pos_filter_data,
                    synth+10,
                    SUBFRAME_SIZE);

            /* Calculate gain of filtered signal for use in AGC. */
            gain_after = 0;
            for (j = 0; j < SUBFRAME_SIZE; j++)
                gain_after += FFABS(synth[j+10]);

            ctx->gain_coeff = ff_g729_adaptive_gain_control(
                    gain_before,
                    gain_after,
                    synth+10,
                    SUBFRAME_SIZE,
                    ctx->gain_coeff);

            if (frame_erasure) {
                ctx->pitch_delay_int_prev = FFMIN(ctx->pitch_delay_int_prev + 1, PITCH_DELAY_MAX);
            } else {
                ctx->pitch_delay_int_prev = pitch_delay_int[i];
            }

            memcpy(synth+8, ctx->hpf_z, 2*sizeof(int16_t));
            ff_acelp_high_pass_filter(
                    out_frame + i*SUBFRAME_SIZE,
                    ctx->hpf_f,
                    synth+10,
                    SUBFRAME_SIZE);
            memcpy(ctx->hpf_z, synth+8+SUBFRAME_SIZE, 2*sizeof(int16_t));
        }

        ctx->was_periodic = is_periodic;

        /* Save signal for use in next frame. */
        memmove(ctx->exc_base, ctx->exc_base + 2 * SUBFRAME_SIZE, (PITCH_DELAY_MAX+INTERPOL_LEN)*sizeof(int16_t));

        buf += format->block_size;
        ctx++;
    }

    *got_frame_ptr = 1;
    return (format->block_size + (avctx->codec_id == AV_CODEC_ID_ACELP_KELVIN)) * avctx->channels;
}

static av_cold int decode_close(AVCodecContext *avctx)
{
    G729Context *s = avctx->priv_data;
    av_freep(&s->channel_context);

    return 0;
}

AVCodec ff_g729_decoder = {
    .name           = "g729",
    .long_name      = NULL_IF_CONFIG_SMALL("G.729"),
    .type           = AVMEDIA_TYPE_AUDIO,
    .id             = AV_CODEC_ID_G729,
    .priv_data_size = sizeof(G729Context),
    .init           = decoder_init,
    .decode         = decode_frame,
    .close          = decode_close,
    .capabilities   = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1,
};

AVCodec ff_acelp_kelvin_decoder = {
    .name           = "acelp.kelvin",
    .long_name      = NULL_IF_CONFIG_SMALL("Sipro ACELP.KELVIN"),
    .type           = AVMEDIA_TYPE_AUDIO,
    .id             = AV_CODEC_ID_ACELP_KELVIN,
    .priv_data_size = sizeof(G729Context),
    .init           = decoder_init,
    .decode         = decode_frame,
    .close          = decode_close,
    .capabilities   = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1,
};