ac3enc.c

/*
 * The simplest AC-3 encoder
 * Copyright (c) 2000 Fabrice Bellard
 *
 * 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
 */

/**
 * @file
 * The simplest AC-3 encoder.
 */

//#define DEBUG

#include "libavcore/audioconvert.h"
#include "libavutil/crc.h"
#include "avcodec.h"
#include "put_bits.h"
#include "ac3.h"
#include "audioconvert.h"


#define MDCT_NBITS 9
#define MDCT_SAMPLES (1 << MDCT_NBITS)

/** Maximum number of exponent groups. +1 for separate DC exponent. */
#define AC3_MAX_EXP_GROUPS 85

/** Scale a float value by 2^bits and convert to an integer. */
#define SCALE_FLOAT(a, bits) lrintf((a) * (float)(1 << (bits)))

/** Scale a float value by 2^15, convert to an integer, and clip to int16_t range. */
#define FIX15(a) av_clip_int16(SCALE_FLOAT(a, 15))


/**
 * Compex number.
 * Used in fixed-point MDCT calculation.
 */
typedef struct IComplex {
    int16_t re,im;
} IComplex;

/**
 * AC-3 encoder private context.
 */
typedef struct AC3EncodeContext {
    PutBitContext pb;                       ///< bitstream writer context

    int bitstream_id;                       ///< bitstream id                           (bsid)
    int bitstream_mode;                     ///< bitstream mode                         (bsmod)

    int bit_rate;                           ///< target bit rate, in bits-per-second
    int sample_rate;                        ///< sampling frequency, in Hz

    int frame_size_min;                     ///< minimum frame size in case rounding is necessary
    int frame_size;                         ///< current frame size in bytes
    int frame_size_code;                    ///< frame size code                        (frmsizecod)
    int bits_written;                       ///< bit count    (used to avg. bitrate)
    int samples_written;                    ///< sample count (used to avg. bitrate)

    int fbw_channels;                       ///< number of full-bandwidth channels      (nfchans)
    int channels;                           ///< total number of channels               (nchans)
    int lfe_on;                             ///< indicates if there is an LFE channel   (lfeon)
    int lfe_channel;                        ///< channel index of the LFE channel
    int channel_mode;                       ///< channel mode                           (acmod)
    const uint8_t *channel_map;             ///< channel map used to reorder channels

    int bandwidth_code[AC3_MAX_CHANNELS];   ///< bandwidth code (0 to 60)               (chbwcod)
    int nb_coefs[AC3_MAX_CHANNELS];

    /* bitrate allocation control */
    int slow_gain_code;                     ///< slow gain code                         (sgaincod)
    int slow_decay_code;                    ///< slow decay code                        (sdcycod)
    int fast_decay_code;                    ///< fast decay code                        (fdcycod)
    int db_per_bit_code;                    ///< dB/bit code                            (dbpbcod)
    int floor_code;                         ///< floor code                             (floorcod)
    AC3BitAllocParameters bit_alloc;        ///< bit allocation parameters
    int coarse_snr_offset;                  ///< coarse SNR offsets                     (csnroffst)
    int fast_gain_code[AC3_MAX_CHANNELS];   ///< fast gain codes (signal-to-mask ratio) (fgaincod)
    int fine_snr_offset[AC3_MAX_CHANNELS];  ///< fine SNR offsets                       (fsnroffst)
    int frame_bits;                         ///< all frame bits except exponents and mantissas
    int exponent_bits;                      ///< number of bits used for exponents

    /* mantissa encoding */
    int mant1_cnt, mant2_cnt, mant4_cnt;    ///< mantissa counts for bap=1,2,4
    uint16_t *qmant1_ptr, *qmant2_ptr, *qmant4_ptr; ///< mantissa pointers for bap=1,2,4

    int16_t last_samples[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE]; ///< last 256 samples from previous frame
} AC3EncodeContext;


/** MDCT and FFT tables */
static int16_t costab[64];
static int16_t sintab[64];
static int16_t xcos1[128];
static int16_t xsin1[128];


/**
 * Adjust the frame size to make the average bit rate match the target bit rate.
 * This is only needed for 11025, 22050, and 44100 sample rates.
 */
static void adjust_frame_size(AC3EncodeContext *s)
{
    while (s->bits_written >= s->bit_rate && s->samples_written >= s->sample_rate) {
        s->bits_written    -= s->bit_rate;
        s->samples_written -= s->sample_rate;
    }
    s->frame_size = s->frame_size_min + 2 * (s->bits_written * s->sample_rate < s->samples_written * s->bit_rate);
    s->bits_written    += s->frame_size * 8;
    s->samples_written += AC3_FRAME_SIZE;
}


/**
 * Deinterleave input samples.
 * Channels are reordered from FFmpeg's default order to AC-3 order.
 */
static void deinterleave_input_samples(AC3EncodeContext *s,
                                       const int16_t *samples,
                                       int16_t planar_samples[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE+AC3_FRAME_SIZE])
{
    int ch, i;

    /* deinterleave and remap input samples */
    for (ch = 0; ch < s->channels; ch++) {
        const int16_t *sptr;
        int sinc;

        /* copy last 256 samples of previous frame to the start of the current frame */
        memcpy(&planar_samples[ch][0], s->last_samples[ch],
               AC3_BLOCK_SIZE * sizeof(planar_samples[0][0]));

        /* deinterleave */
        sinc = s->channels;
        sptr = samples + s->channel_map[ch];
        for (i = AC3_BLOCK_SIZE; i < AC3_FRAME_SIZE+AC3_BLOCK_SIZE; i++) {
            planar_samples[ch][i] = *sptr;
            sptr += sinc;
        }

        /* save last 256 samples for next frame */
        memcpy(s->last_samples[ch], &planar_samples[ch][6* AC3_BLOCK_SIZE],
               AC3_BLOCK_SIZE * sizeof(planar_samples[0][0]));
    }
}


/**
 * Initialize FFT tables.
 * @param ln log2(FFT size)
 */
static av_cold void fft_init(int ln)
{
    int i, n, n2;
    float alpha;

    n  = 1 << ln;
    n2 = n >> 1;

    for (i = 0; i < n2; i++) {
        alpha     = 2.0 * M_PI * i / n;
        costab[i] = FIX15(cos(alpha));
        sintab[i] = FIX15(sin(alpha));
    }
}


/**
 * Initialize MDCT tables.
 * @param nbits log2(MDCT size)
 */
static av_cold void mdct_init(int nbits)
{
    int i, n, n4;

    n  = 1 << nbits;
    n4 = n >> 2;

    fft_init(nbits - 2);

    for (i = 0; i < n4; i++) {
        float alpha = 2.0 * M_PI * (i + 1.0 / 8.0) / n;
        xcos1[i] = FIX15(-cos(alpha));
        xsin1[i] = FIX15(-sin(alpha));
    }
}


/** Butterfly op */
#define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1)  \
{                                                       \
  int ax, ay, bx, by;                                   \
  bx  = pre1;                                           \
  by  = pim1;                                           \
  ax  = qre1;                                           \
  ay  = qim1;                                           \
  pre = (bx + ax) >> 1;                                 \
  pim = (by + ay) >> 1;                                 \
  qre = (bx - ax) >> 1;                                 \
  qim = (by - ay) >> 1;                                 \
}


/** Complex multiply */
#define CMUL(pre, pim, are, aim, bre, bim)              \
{                                                       \
   pre = (MUL16(are, bre) - MUL16(aim, bim)) >> 15;     \
   pim = (MUL16(are, bim) + MUL16(bre, aim)) >> 15;     \
}


/**
 * Calculate a 2^n point complex FFT on 2^ln points.
 * @param z  complex input/output samples
 * @param ln log2(FFT size)
 */
static void fft(IComplex *z, int ln)
{
    int j, l, np, np2;
    int nblocks, nloops;
    register IComplex *p,*q;
    int tmp_re, tmp_im;

    np = 1 << ln;

    /* reverse */
    for (j = 0; j < np; j++) {
        int k = av_reverse[j] >> (8 - ln);
        if (k < j)
            FFSWAP(IComplex, z[k], z[j]);
    }

    /* pass 0 */

    p = &z[0];
    j = np >> 1;
    do {
        BF(p[0].re, p[0].im, p[1].re, p[1].im,
           p[0].re, p[0].im, p[1].re, p[1].im);
        p += 2;
    } while (--j);

    /* pass 1 */

    p = &z[0];
    j = np >> 2;
    do {
        BF(p[0].re, p[0].im, p[2].re,  p[2].im,
           p[0].re, p[0].im, p[2].re,  p[2].im);
        BF(p[1].re, p[1].im, p[3].re,  p[3].im,
           p[1].re, p[1].im, p[3].im, -p[3].re);
        p+=4;
    } while (--j);

    /* pass 2 .. ln-1 */

    nblocks = np >> 3;
    nloops  =  1 << 2;
    np2     = np >> 1;
    do {
        p = z;
        q = z + nloops;
        for (j = 0; j < nblocks; j++) {
            BF(p->re, p->im, q->re, q->im,
               p->re, p->im, q->re, q->im);
            p++;
            q++;
            for(l = nblocks; l < np2; l += nblocks) {
                CMUL(tmp_re, tmp_im, costab[l], -sintab[l], q->re, q->im);
                BF(p->re, p->im, q->re,  q->im,
                   p->re, p->im, tmp_re, tmp_im);
                p++;
                q++;
            }
            p += nloops;
            q += nloops;
        }
        nblocks = nblocks >> 1;
        nloops  = nloops  << 1;
    } while (nblocks);
}


/**
 * Calculate a 512-point MDCT
 * @param out 256 output frequency coefficients
 * @param in  512 windowed input audio samples
 */
static void mdct512(int32_t *out, int16_t *in)
{
    int i, re, im, re1, im1;
    int16_t rot[MDCT_SAMPLES];
    IComplex x[MDCT_SAMPLES/4];

    /* shift to simplify computations */
    for (i = 0; i < MDCT_SAMPLES/4; i++)
        rot[i] = -in[i + 3*MDCT_SAMPLES/4];
    for (;i < MDCT_SAMPLES; i++)
        rot[i] =  in[i -   MDCT_SAMPLES/4];

    /* pre rotation */
    for (i = 0; i < MDCT_SAMPLES/4; i++) {
        re =  ((int)rot[               2*i] - (int)rot[MDCT_SAMPLES  -1-2*i]) >> 1;
        im = -((int)rot[MDCT_SAMPLES/2+2*i] - (int)rot[MDCT_SAMPLES/2-1-2*i]) >> 1;
        CMUL(x[i].re, x[i].im, re, im, -xcos1[i], xsin1[i]);
    }

    fft(x, MDCT_NBITS - 2);

    /* post rotation */
    for (i = 0; i < MDCT_SAMPLES/4; i++) {
        re = x[i].re;
        im = x[i].im;
        CMUL(re1, im1, re, im, xsin1[i], xcos1[i]);
        out[                 2*i] = im1;
        out[MDCT_SAMPLES/2-1-2*i] = re1;
    }
}


/**
 * Apply KBD window to input samples prior to MDCT.
 */
static void apply_window(int16_t *output, const int16_t *input,
                         const int16_t *window, int n)
{
    int i;
    int n2 = n >> 1;

    for (i = 0; i < n2; i++) {
        output[i]     = MUL16(input[i],     window[i]) >> 15;
        output[n-i-1] = MUL16(input[n-i-1], window[i]) >> 15;
    }
}


/**
 * Calculate the log2() of the maximum absolute value in an array.
 * @param tab input array
 * @param n   number of values in the array
 * @return    log2(max(abs(tab[])))
 */
static int log2_tab(int16_t *tab, int n)
{
    int i, v;

    v = 0;
    for (i = 0; i < n; i++)
        v |= abs(tab[i]);

    return av_log2(v);
}


/**
 * Left-shift each value in an array by a specified amount.
 * @param tab    input array
 * @param n      number of values in the array
 * @param lshift left shift amount. a negative value means right shift.
 */
static void lshift_tab(int16_t *tab, int n, int lshift)
{
    int i;

    if (lshift > 0) {
        for(i = 0; i < n; i++)
            tab[i] <<= lshift;
    } else if (lshift < 0) {
        lshift = -lshift;
        for (i = 0; i < n; i++)
            tab[i] >>= lshift;
    }
}


/**
 * Normalize the input samples to use the maximum available precision.
 * This assumes signed 16-bit input samples. Exponents are reduced by 9 to
 * match the 24-bit internal precision for MDCT coefficients.
 *
 * @return exponent shift
 */
static int normalize_samples(AC3EncodeContext *s,
                             int16_t windowed_samples[AC3_WINDOW_SIZE])
{
    int v = 14 - log2_tab(windowed_samples, AC3_WINDOW_SIZE);
    v = FFMAX(0, v);
    lshift_tab(windowed_samples, AC3_WINDOW_SIZE, v);
    return v - 9;
}


/**
 * Apply the MDCT to input samples to generate frequency coefficients.
 * This applies the KBD window and normalizes the input to reduce precision
 * loss due to fixed-point calculations.
 */
static void apply_mdct(AC3EncodeContext *s,
                       int16_t planar_samples[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE+AC3_FRAME_SIZE],
                       int8_t exp_shift[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                       int32_t mdct_coef[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS])
{
    int blk, ch;
    int16_t windowed_samples[AC3_WINDOW_SIZE];

    for (ch = 0; ch < s->channels; ch++) {
        for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
            const int16_t *input_samples = &planar_samples[ch][blk * AC3_BLOCK_SIZE];

            apply_window(windowed_samples, input_samples, ff_ac3_window, AC3_WINDOW_SIZE);

            exp_shift[blk][ch] = normalize_samples(s, windowed_samples);

            mdct512(mdct_coef[blk][ch], windowed_samples);
        }
    }
}


/**
 * Extract exponents from the MDCT coefficients.
 * This takes into account the normalization that was done to the input samples
 * by adjusting the exponents by the exponent shift values.
 */
static void extract_exponents(AC3EncodeContext *s,
                              int32_t mdct_coef[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                              int8_t exp_shift[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                              uint8_t exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS])
{
    int blk, ch, i;

    /* extract exponents */
    for (ch = 0; ch < s->channels; ch++) {
        for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
            /* compute "exponents". We take into account the normalization there */
            for (i = 0; i < AC3_MAX_COEFS; i++) {
                int e;
                int v = abs(mdct_coef[blk][ch][i]);
                if (v == 0)
                    e = 24;
                else {
                    e = 23 - av_log2(v) + exp_shift[blk][ch];
                    if (e >= 24) {
                        e = 24;
                        mdct_coef[blk][ch][i] = 0;
                    }
                }
                exp[blk][ch][i] = e;
            }
        }
    }
}


/**
 * Calculate the sum of absolute differences (SAD) between 2 sets of exponents.
 */
static int calc_exp_diff(uint8_t *exp1, uint8_t *exp2, int n)
{
    int sum, i;
    sum = 0;
    for (i = 0; i < n; i++)
        sum += abs(exp1[i] - exp2[i]);
    return sum;
}


/**
 * Exponent Difference Threshold.
 * New exponents are sent if their SAD exceed this number.
 */
#define EXP_DIFF_THRESHOLD 1000


/**
 * Calculate exponent strategies for all blocks in a single channel.
 */
static void compute_exp_strategy_ch(uint8_t *exp_strategy, uint8_t **exp)
{
    int blk, blk1;
    int exp_diff;

    /* estimate if the exponent variation & decide if they should be
       reused in the next frame */
    exp_strategy[0] = EXP_NEW;
    for (blk = 1; blk < AC3_MAX_BLOCKS; blk++) {
        exp_diff = calc_exp_diff(exp[blk], exp[blk-1], AC3_MAX_COEFS);
        if (exp_diff > EXP_DIFF_THRESHOLD)
            exp_strategy[blk] = EXP_NEW;
        else
            exp_strategy[blk] = EXP_REUSE;
    }

    /* now select the encoding strategy type : if exponents are often
       recoded, we use a coarse encoding */
    blk = 0;
    while (blk < AC3_MAX_BLOCKS) {
        blk1 = blk + 1;
        while (blk1 < AC3_MAX_BLOCKS && exp_strategy[blk1] == EXP_REUSE)
            blk1++;
        switch (blk1 - blk) {
        case 1:  exp_strategy[blk] = EXP_D45; break;
        case 2:
        case 3:  exp_strategy[blk] = EXP_D25; break;
        default: exp_strategy[blk] = EXP_D15; break;
        }
        blk = blk1;
    }
}


/**
 * Calculate exponent strategies for all channels.
 * Array arrangement is reversed to simplify the per-channel calculation.
 */
static void compute_exp_strategy(AC3EncodeContext *s,
                                 uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                                 uint8_t exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS])
{
    uint8_t *exp1[AC3_MAX_CHANNELS][AC3_MAX_BLOCKS];
    uint8_t exp_str1[AC3_MAX_CHANNELS][AC3_MAX_BLOCKS];
    int ch, blk;

    for (ch = 0; ch < s->fbw_channels; ch++) {
        for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
            exp1[ch][blk]     = exp[blk][ch];
            exp_str1[ch][blk] = exp_strategy[blk][ch];
        }

        compute_exp_strategy_ch(exp_str1[ch], exp1[ch]);

        for (blk = 0; blk < AC3_MAX_BLOCKS; blk++)
            exp_strategy[blk][ch] = exp_str1[ch][blk];
    }
    if (s->lfe_on) {
        ch = s->lfe_channel;
        exp_strategy[0][ch] = EXP_D15;
        for (blk = 1; blk < 5; blk++)
            exp_strategy[blk][ch] = EXP_REUSE;
    }
}


/**
 * Set each encoded exponent in a block to the minimum of itself and the
 * exponent in the same frequency bin of a following block.
 * exp[i] = min(exp[i], exp1[i]
 */
static void exponent_min(uint8_t exp[AC3_MAX_COEFS], uint8_t exp1[AC3_MAX_COEFS], int n)
{
    int i;
    for (i = 0; i < n; i++) {
        if (exp1[i] < exp[i])
            exp[i] = exp1[i];
    }
}


/**
 * Update the exponents so that they are the ones the decoder will decode.
 */
static void encode_exponents_blk_ch(uint8_t encoded_exp[AC3_MAX_COEFS],
                                    uint8_t exp[AC3_MAX_COEFS],
                                    int nb_exps, int exp_strategy,
                                    uint8_t *num_exp_groups)
{
    int group_size, nb_groups, i, j, k, exp_min;
    uint8_t exp1[AC3_MAX_COEFS];

    group_size = exp_strategy + (exp_strategy == EXP_D45);
    *num_exp_groups = (nb_exps + (group_size * 3) - 4) / (3 * group_size);
    nb_groups = *num_exp_groups * 3;

    /* for each group, compute the minimum exponent */
    exp1[0] = exp[0]; /* DC exponent is handled separately */
    k = 1;
    for (i = 1; i <= nb_groups; i++) {
        exp_min = exp[k];
        assert(exp_min >= 0 && exp_min <= 24);
        for (j = 1; j < group_size; j++) {
            if (exp[k+j] < exp_min)
                exp_min = exp[k+j];
        }
        exp1[i] = exp_min;
        k += group_size;
    }

    /* constraint for DC exponent */
    if (exp1[0] > 15)
        exp1[0] = 15;

    /* decrease the delta between each groups to within 2 so that they can be
       differentially encoded */
    for (i = 1; i <= nb_groups; i++)
        exp1[i] = FFMIN(exp1[i], exp1[i-1] + 2);
    for (i = nb_groups-1; i >= 0; i--)
        exp1[i] = FFMIN(exp1[i], exp1[i+1] + 2);

    /* now we have the exponent values the decoder will see */
    encoded_exp[0] = exp1[0];
    k = 1;
    for (i = 1; i <= nb_groups; i++) {
        for (j = 0; j < group_size; j++)
            encoded_exp[k+j] = exp1[i];
        k += group_size;
    }
}


/**
 * Encode exponents from original extracted form to what the decoder will see.
 * This copies and groups exponents based on exponent strategy and reduces
 * deltas between adjacent exponent groups so that they can be differentially
 * encoded.
 */
static void encode_exponents(AC3EncodeContext *s,
                             uint8_t exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                             uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                             uint8_t num_exp_groups[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                             uint8_t encoded_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS])
{
    int blk, blk1, blk2, ch;

    for (ch = 0; ch < s->channels; ch++) {
        /* for the EXP_REUSE case we select the min of the exponents */
        blk = 0;
        while (blk < AC3_MAX_BLOCKS) {
            blk1 = blk + 1;
            while (blk1 < AC3_MAX_BLOCKS && exp_strategy[blk1][ch] == EXP_REUSE) {
                exponent_min(exp[blk][ch], exp[blk1][ch], s->nb_coefs[ch]);
                blk1++;
            }
            encode_exponents_blk_ch(encoded_exp[blk][ch],
                                    exp[blk][ch], s->nb_coefs[ch],
                                    exp_strategy[blk][ch],
                                    &num_exp_groups[blk][ch]);
            /* copy encoded exponents for reuse case */
            for (blk2 = blk+1; blk2 < blk1; blk2++) {
                memcpy(encoded_exp[blk2][ch], encoded_exp[blk][ch],
                       s->nb_coefs[ch] * sizeof(uint8_t));
            }
            blk = blk1;
        }
    }
}


/**
 * Group exponents.
 * 3 delta-encoded exponents are in each 7-bit group. The number of groups
 * varies depending on exponent strategy and bandwidth.
 */
static void group_exponents(AC3EncodeContext *s,
                            uint8_t encoded_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                            uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                            uint8_t num_exp_groups[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                            uint8_t grouped_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_EXP_GROUPS])
{
    int blk, ch, i;
    int group_size, bit_count;
    uint8_t *p;
    int delta0, delta1, delta2;
    int exp0, exp1;

    bit_count = 0;
    for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
        for (ch = 0; ch < s->channels; ch++) {
            if (exp_strategy[blk][ch] == EXP_REUSE) {
                num_exp_groups[blk][ch] = 0;
                continue;
            }
            group_size = exp_strategy[blk][ch] + (exp_strategy[blk][ch] == EXP_D45);
            bit_count += 4 + (num_exp_groups[blk][ch] * 7);
            p = encoded_exp[blk][ch];

            /* DC exponent */
            exp1 = *p++;
            grouped_exp[blk][ch][0] = exp1;

            /* remaining exponents are delta encoded */
            for (i = 1; i <= num_exp_groups[blk][ch]; i++) {
                /* merge three delta in one code */
                exp0   = exp1;
                exp1   = p[0];
                p     += group_size;
                delta0 = exp1 - exp0 + 2;

                exp0   = exp1;
                exp1   = p[0];
                p     += group_size;
                delta1 = exp1 - exp0 + 2;

                exp0   = exp1;
                exp1   = p[0];
                p     += group_size;
                delta2 = exp1 - exp0 + 2;

                grouped_exp[blk][ch][i] = ((delta0 * 5 + delta1) * 5) + delta2;
            }
        }
    }

    s->exponent_bits = bit_count;
}


/**
 * Calculate final exponents from the supplied MDCT coefficients and exponent shift.
 * Extract exponents from MDCT coefficients, calculate exponent strategies,
 * and encode final exponents.
 */
static void process_exponents(AC3EncodeContext *s,
                              int32_t mdct_coef[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                              int8_t exp_shift[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                              uint8_t exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                              uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                              uint8_t encoded_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                              uint8_t num_exp_groups[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                              uint8_t grouped_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_EXP_GROUPS])
{
    extract_exponents(s, mdct_coef, exp_shift, exp);

    compute_exp_strategy(s, exp_strategy, exp);

    encode_exponents(s, exp, exp_strategy, num_exp_groups, encoded_exp);

    group_exponents(s, encoded_exp, exp_strategy, num_exp_groups, grouped_exp);
}


/**
 * Initialize bit allocation.
 * Set default parameter codes and calculate parameter values.
 */
static void bit_alloc_init(AC3EncodeContext *s)
{
    int ch;

    /* init default parameters */
    s->slow_decay_code = 2;
    s->fast_decay_code = 1;
    s->slow_gain_code  = 1;
    s->db_per_bit_code = 2;
    s->floor_code      = 4;
    for (ch = 0; ch < s->channels; ch++)
        s->fast_gain_code[ch] = 4;

    /* initial snr offset */
    s->coarse_snr_offset = 40;

    /* compute real values */
    /* currently none of these values change during encoding, so we can just
       set them once at initialization */
    s->bit_alloc.slow_decay = ff_ac3_slow_decay_tab[s->slow_decay_code] >> s->bit_alloc.sr_shift;
    s->bit_alloc.fast_decay = ff_ac3_fast_decay_tab[s->fast_decay_code] >> s->bit_alloc.sr_shift;
    s->bit_alloc.slow_gain  = ff_ac3_slow_gain_tab[s->slow_gain_code];
    s->bit_alloc.db_per_bit = ff_ac3_db_per_bit_tab[s->db_per_bit_code];
    s->bit_alloc.floor      = ff_ac3_floor_tab[s->floor_code];
}


/**
 * Count the bits used to encode the frame, minus exponents and mantissas.
 */
static void count_frame_bits(AC3EncodeContext *s,
                             uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS])
{
    static const int frame_bits_inc[8] = { 0, 0, 2, 2, 2, 4, 2, 4 };
    int blk, ch;
    int frame_bits;

    /* header size */
    frame_bits = 65;
    frame_bits += frame_bits_inc[s->channel_mode];

    /* audio blocks */
    for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
        frame_bits += s->fbw_channels * 2 + 2; /* blksw * c, dithflag * c, dynrnge, cplstre */
        if (s->channel_mode == AC3_CHMODE_STEREO) {
            frame_bits++; /* rematstr */
            if (!blk)
                frame_bits += 4;
        }
        frame_bits += 2 * s->fbw_channels; /* chexpstr[2] * c */
        if (s->lfe_on)
            frame_bits++; /* lfeexpstr */
        for (ch = 0; ch < s->fbw_channels; ch++) {
            if (exp_strategy[blk][ch] != EXP_REUSE)
                frame_bits += 6 + 2; /* chbwcod[6], gainrng[2] */
        }
        frame_bits++; /* baie */
        frame_bits++; /* snr */
        frame_bits += 2; /* delta / skip */
    }
    frame_bits++; /* cplinu for block 0 */
    /* bit alloc info */
    /* sdcycod[2], fdcycod[2], sgaincod[2], dbpbcod[2], floorcod[3] */
    /* csnroffset[6] */
    /* (fsnoffset[4] + fgaincod[4]) * c */
    frame_bits += 2*4 + 3 + 6 + s->channels * (4 + 3);

    /* auxdatae, crcrsv */
    frame_bits += 2;

    /* CRC */
    frame_bits += 16;

    s->frame_bits = frame_bits;
}


/**
 * Calculate the number of bits needed to encode a set of mantissas.
 */
static int compute_mantissa_size(AC3EncodeContext *s, uint8_t *m, int nb_coefs)
{
    int bits, mant, i;

    bits = 0;
    for (i = 0; i < nb_coefs; i++) {
        mant = m[i];
        switch (mant) {
        case 0:
            /* nothing */
            break;
        case 1:
            /* 3 mantissa in 5 bits */
            if (s->mant1_cnt == 0)
                bits += 5;
            if (++s->mant1_cnt == 3)
                s->mant1_cnt = 0;
            break;
        case 2:
            /* 3 mantissa in 7 bits */
            if (s->mant2_cnt == 0)
                bits += 7;
            if (++s->mant2_cnt == 3)
                s->mant2_cnt = 0;
            break;
        case 3:
            bits += 3;
            break;
        case 4:
            /* 2 mantissa in 7 bits */
            if (s->mant4_cnt == 0)
                bits += 7;
            if (++s->mant4_cnt == 2)
                s->mant4_cnt = 0;
            break;
        case 14:
            bits += 14;
            break;
        case 15:
            bits += 16;
            break;
        default:
            bits += mant - 1;
            break;
        }
    }
    return bits;
}


/**
 * Calculate masking curve based on the final exponents.
 * Also calculate the power spectral densities to use in future calculations.
 */
static void bit_alloc_masking(AC3EncodeContext *s,
                              uint8_t encoded_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                              uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
                              int16_t psd[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                              int16_t mask[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_CRITICAL_BANDS])
{
    int blk, ch;
    int16_t band_psd[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_CRITICAL_BANDS];

    for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
        for (ch = 0; ch < s->channels; ch++) {
            if(exp_strategy[blk][ch] == EXP_REUSE) {
                memcpy(psd[blk][ch],  psd[blk-1][ch],  AC3_MAX_COEFS*sizeof(psd[0][0][0]));
                memcpy(mask[blk][ch], mask[blk-1][ch], AC3_CRITICAL_BANDS*sizeof(mask[0][0][0]));
            } else {
                ff_ac3_bit_alloc_calc_psd(encoded_exp[blk][ch], 0,
                                          s->nb_coefs[ch],
                                          psd[blk][ch], band_psd[blk][ch]);
                ff_ac3_bit_alloc_calc_mask(&s->bit_alloc, band_psd[blk][ch],
                                           0, s->nb_coefs[ch],
                                           ff_ac3_fast_gain_tab[s->fast_gain_code[ch]],
                                           ch == s->lfe_channel,
                                           DBA_NONE, 0, NULL, NULL, NULL,
                                           mask[blk][ch]);
            }
        }
    }
}


/**
 * Run the bit allocation with a given SNR offset.
 * This calculates the bit allocation pointers that will be used to determine
 * the quantization of each mantissa.
 * @return the number of bits needed for mantissas if the given SNR offset is
 *         is used.
 */
static int bit_alloc(AC3EncodeContext *s,
                     int16_t mask[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_CRITICAL_BANDS],
                     int16_t psd[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                     uint8_t bap[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                     int snr_offset)
{
    int blk, ch;
    int mantissa_bits;

    snr_offset = (snr_offset - 240) << 2;

    mantissa_bits = 0;
    for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
        s->mant1_cnt = 0;
        s->mant2_cnt = 0;
        s->mant4_cnt = 0;
        for (ch = 0; ch < s->channels; ch++) {
            ff_ac3_bit_alloc_calc_bap(mask[blk][ch], psd[blk][ch], 0,
                                      s->nb_coefs[ch], snr_offset,
                                      s->bit_alloc.floor, ff_ac3_bap_tab,
                                      bap[blk][ch]);
            mantissa_bits += compute_mantissa_size(s, bap[blk][ch], s->nb_coefs[ch]);
        }
    }
    return mantissa_bits;
}


/**
 * Perform bit allocation search.
 * Finds the SNR offset value that maximizes quality and fits in the specified
 * frame size.  Output is the SNR offset and a set of bit allocation pointers
 * used to quantize the mantissas.
 */
static int compute_bit_allocation(AC3EncodeContext *s,
                                  uint8_t bap[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                                  uint8_t encoded_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
                                  uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS])
{
    int ch;
    int bits_left;
    int snr_offset;
    uint8_t bap1[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS];
    int16_t psd[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS];
    int16_t mask[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_CRITICAL_BANDS];

    /* count frame bits other than exponents and mantissas */
    count_frame_bits(s, exp_strategy);

    /* calculate psd and masking curve before doing bit allocation */
    bit_alloc_masking(s, encoded_exp, exp_strategy, psd, mask);

    /* now the big work begins : do the bit allocation. Modify the snr
       offset until we can pack everything in the requested frame size */

    bits_left = 8 * s->frame_size - (s->frame_bits + s->exponent_bits);

    snr_offset = s->coarse_snr_offset << 4;

    while (snr_offset >= 0 &&
           bit_alloc(s, mask, psd, bap, snr_offset) > bits_left) {
        snr_offset -= 64;
    }
    if (snr_offset < 0) {
        return AVERROR(EINVAL);
    }

    while (snr_offset + 64 <= 1023 &&
           bit_alloc(s, mask, psd, bap1, snr_offset + 64) <= bits_left) {
        snr_offset += 64;
        memcpy(bap, bap1, sizeof(bap1));
    }
    while (snr_offset + 16 <= 1023 &&
           bit_alloc(s, mask, psd, bap1, snr_offset + 16) <= bits_left) {
        snr_offset += 16;
        memcpy(bap, bap1, sizeof(bap1));
    }
    while (snr_offset + 4 <= 1023 &&
           bit_alloc(s, mask, psd, bap1, snr_offset + 4) <= bits_left) {