mpegaudiodec_template.c

/*
 * MPEG Audio decoder
 * Copyright (c) 2001, 2002 Fabrice Bellard
 *
 * This file is part of Libav.
 *
 * Libav 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.
 *
 * Libav 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 Libav; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @file
 * MPEG Audio decoder
 */

#include "libavutil/attributes.h"
#include "libavutil/avassert.h"
#include "libavutil/channel_layout.h"
#include "libavutil/float_dsp.h"
#include "avcodec.h"
#include "get_bits.h"
#include "internal.h"
#include "mathops.h"
#include "mpegaudiodsp.h"

/*
 * TODO:
 *  - test lsf / mpeg25 extensively.
 */

#include "mpegaudio.h"
#include "mpegaudiodecheader.h"

#define BACKSTEP_SIZE 512
#define EXTRABYTES 24
#define LAST_BUF_SIZE 2 * BACKSTEP_SIZE + EXTRABYTES

/* layer 3 "granule" */
typedef struct GranuleDef {
    uint8_t scfsi;
    int part2_3_length;
    int big_values;
    int global_gain;
    int scalefac_compress;
    uint8_t block_type;
    uint8_t switch_point;
    int table_select[3];
    int subblock_gain[3];
    uint8_t scalefac_scale;
    uint8_t count1table_select;
    int region_size[3]; /* number of huffman codes in each region */
    int preflag;
    int short_start, long_end; /* long/short band indexes */
    uint8_t scale_factors[40];
    DECLARE_ALIGNED(16, INTFLOAT, sb_hybrid)[SBLIMIT * 18]; /* 576 samples */
} GranuleDef;

typedef struct MPADecodeContext {
    MPA_DECODE_HEADER
    uint8_t last_buf[LAST_BUF_SIZE];
    int last_buf_size;
    /* next header (used in free format parsing) */
    uint32_t free_format_next_header;
    GetBitContext gb;
    GetBitContext in_gb;
    DECLARE_ALIGNED(32, MPA_INT, synth_buf)[MPA_MAX_CHANNELS][512 * 2];
    int synth_buf_offset[MPA_MAX_CHANNELS];
    DECLARE_ALIGNED(32, INTFLOAT, sb_samples)[MPA_MAX_CHANNELS][36][SBLIMIT];
    INTFLOAT mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
    GranuleDef granules[2][2]; /* Used in Layer 3 */
    int adu_mode; ///< 0 for standard mp3, 1 for adu formatted mp3
    int dither_state;
    int err_recognition;
    AVCodecContext* avctx;
    MPADSPContext mpadsp;
    AVFloatDSPContext fdsp;
    AVFrame *frame;
} MPADecodeContext;

#define HEADER_SIZE 4

#include "mpegaudiodata.h"
#include "mpegaudiodectab.h"

/* vlc structure for decoding layer 3 huffman tables */
static VLC huff_vlc[16];
static VLC_TYPE huff_vlc_tables[
    0 + 128 + 128 + 128 + 130 + 128 + 154 + 166 +
  142 + 204 + 190 + 170 + 542 + 460 + 662 + 414
  ][2];
static const int huff_vlc_tables_sizes[16] = {
    0,  128,  128,  128,  130,  128,  154,  166,
  142,  204,  190,  170,  542,  460,  662,  414
};
static VLC huff_quad_vlc[2];
static VLC_TYPE  huff_quad_vlc_tables[128+16][2];
static const int huff_quad_vlc_tables_sizes[2] = { 128, 16 };
/* computed from band_size_long */
static uint16_t band_index_long[9][23];
#include "mpegaudio_tablegen.h"
/* intensity stereo coef table */
static INTFLOAT is_table[2][16];
static INTFLOAT is_table_lsf[2][2][16];
static INTFLOAT csa_table[8][4];

static int16_t division_tab3[1<<6 ];
static int16_t division_tab5[1<<8 ];
static int16_t division_tab9[1<<11];

static int16_t * const division_tabs[4] = {
    division_tab3, division_tab5, NULL, division_tab9
};

/* lower 2 bits: modulo 3, higher bits: shift */
static uint16_t scale_factor_modshift[64];
/* [i][j]:  2^(-j/3) * FRAC_ONE * 2^(i+2) / (2^(i+2) - 1) */
static int32_t scale_factor_mult[15][3];
/* mult table for layer 2 group quantization */

#define SCALE_GEN(v) \
{ FIXR_OLD(1.0 * (v)), FIXR_OLD(0.7937005259 * (v)), FIXR_OLD(0.6299605249 * (v)) }

static const int32_t scale_factor_mult2[3][3] = {
    SCALE_GEN(4.0 / 3.0), /* 3 steps */
    SCALE_GEN(4.0 / 5.0), /* 5 steps */
    SCALE_GEN(4.0 / 9.0), /* 9 steps */
};

/**
 * Convert region offsets to region sizes and truncate
 * size to big_values.
 */
static void region_offset2size(GranuleDef *g)
{
    int i, k, j = 0;
    g->region_size[2] = 576 / 2;
    for (i = 0; i < 3; i++) {
        k = FFMIN(g->region_size[i], g->big_values);
        g->region_size[i] = k - j;
        j = k;
    }
}

static void init_short_region(MPADecodeContext *s, GranuleDef *g)
{
    if (g->block_type == 2) {
        if (s->sample_rate_index != 8)
            g->region_size[0] = (36 / 2);
        else
            g->region_size[0] = (72 / 2);
    } else {
        if (s->sample_rate_index <= 2)
            g->region_size[0] = (36 / 2);
        else if (s->sample_rate_index != 8)
            g->region_size[0] = (54 / 2);
        else
            g->region_size[0] = (108 / 2);
    }
    g->region_size[1] = (576 / 2);
}

static void init_long_region(MPADecodeContext *s, GranuleDef *g,
                             int ra1, int ra2)
{
    int l;
    g->region_size[0] = band_index_long[s->sample_rate_index][ra1 + 1] >> 1;
    /* should not overflow */
    l = FFMIN(ra1 + ra2 + 2, 22);
    g->region_size[1] = band_index_long[s->sample_rate_index][      l] >> 1;
}

static void compute_band_indexes(MPADecodeContext *s, GranuleDef *g)
{
    if (g->block_type == 2) {
        if (g->switch_point) {
            /* if switched mode, we handle the 36 first samples as
                long blocks.  For 8000Hz, we handle the 72 first
                exponents as long blocks */
            if (s->sample_rate_index <= 2)
                g->long_end = 8;
            else
                g->long_end = 6;

            g->short_start = 3;
        } else {
            g->long_end    = 0;
            g->short_start = 0;
        }
    } else {
        g->short_start = 13;
        g->long_end    = 22;
    }
}

/* layer 1 unscaling */
/* n = number of bits of the mantissa minus 1 */
static inline int l1_unscale(int n, int mant, int scale_factor)
{
    int shift, mod;
    int64_t val;

    shift   = scale_factor_modshift[scale_factor];
    mod     = shift & 3;
    shift >>= 2;
    val     = MUL64(mant + (-1 << n) + 1, scale_factor_mult[n-1][mod]);
    shift  += n;
    /* NOTE: at this point, 1 <= shift >= 21 + 15 */
    return (int)((val + (1LL << (shift - 1))) >> shift);
}

static inline int l2_unscale_group(int steps, int mant, int scale_factor)
{
    int shift, mod, val;

    shift   = scale_factor_modshift[scale_factor];
    mod     = shift & 3;
    shift >>= 2;

    val = (mant - (steps >> 1)) * scale_factor_mult2[steps >> 2][mod];
    /* NOTE: at this point, 0 <= shift <= 21 */
    if (shift > 0)
        val = (val + (1 << (shift - 1))) >> shift;
    return val;
}

/* compute value^(4/3) * 2^(exponent/4). It normalized to FRAC_BITS */
static inline int l3_unscale(int value, int exponent)
{
    unsigned int m;
    int e;

    e  = table_4_3_exp  [4 * value + (exponent & 3)];
    m  = table_4_3_value[4 * value + (exponent & 3)];
    e -= exponent >> 2;
    assert(e >= 1);
    if (e > 31)
        return 0;
    m = (m + (1 << (e - 1))) >> e;

    return m;
}

static av_cold void decode_init_static(void)
{
    int i, j, k;
    int offset;

    /* scale factors table for layer 1/2 */
    for (i = 0; i < 64; i++) {
        int shift, mod;
        /* 1.0 (i = 3) is normalized to 2 ^ FRAC_BITS */
        shift = i / 3;
        mod   = i % 3;
        scale_factor_modshift[i] = mod | (shift << 2);
    }

    /* scale factor multiply for layer 1 */
    for (i = 0; i < 15; i++) {
        int n, norm;
        n = i + 2;
        norm = ((INT64_C(1) << n) * FRAC_ONE) / ((1 << n) - 1);
        scale_factor_mult[i][0] = MULLx(norm, FIXR(1.0          * 2.0), FRAC_BITS);
        scale_factor_mult[i][1] = MULLx(norm, FIXR(0.7937005259 * 2.0), FRAC_BITS);
        scale_factor_mult[i][2] = MULLx(norm, FIXR(0.6299605249 * 2.0), FRAC_BITS);
        ff_dlog(NULL, "%d: norm=%x s=%x %x %x\n", i, norm,
                scale_factor_mult[i][0],
                scale_factor_mult[i][1],
                scale_factor_mult[i][2]);
    }

    RENAME(ff_mpa_synth_init)(RENAME(ff_mpa_synth_window));

    /* huffman decode tables */
    offset = 0;
    for (i = 1; i < 16; i++) {
        const HuffTable *h = &mpa_huff_tables[i];
        int xsize, x, y;
        uint8_t  tmp_bits [512] = { 0 };
        uint16_t tmp_codes[512] = { 0 };

        xsize = h->xsize;

        j = 0;
        for (x = 0; x < xsize; x++) {
            for (y = 0; y < xsize; y++) {
                tmp_bits [(x << 5) | y | ((x&&y)<<4)]= h->bits [j  ];
                tmp_codes[(x << 5) | y | ((x&&y)<<4)]= h->codes[j++];
            }
        }

        /* XXX: fail test */
        huff_vlc[i].table = huff_vlc_tables+offset;
        huff_vlc[i].table_allocated = huff_vlc_tables_sizes[i];
        init_vlc(&huff_vlc[i], 7, 512,
                 tmp_bits, 1, 1, tmp_codes, 2, 2,
                 INIT_VLC_USE_NEW_STATIC);
        offset += huff_vlc_tables_sizes[i];
    }
    assert(offset == FF_ARRAY_ELEMS(huff_vlc_tables));

    offset = 0;
    for (i = 0; i < 2; i++) {
        huff_quad_vlc[i].table = huff_quad_vlc_tables+offset;
        huff_quad_vlc[i].table_allocated = huff_quad_vlc_tables_sizes[i];
        init_vlc(&huff_quad_vlc[i], i == 0 ? 7 : 4, 16,
                 mpa_quad_bits[i], 1, 1, mpa_quad_codes[i], 1, 1,
                 INIT_VLC_USE_NEW_STATIC);
        offset += huff_quad_vlc_tables_sizes[i];
    }
    assert(offset == FF_ARRAY_ELEMS(huff_quad_vlc_tables));

    for (i = 0; i < 9; i++) {
        k = 0;
        for (j = 0; j < 22; j++) {
            band_index_long[i][j] = k;
            k += band_size_long[i][j];
        }
        band_index_long[i][22] = k;
    }

    /* compute n ^ (4/3) and store it in mantissa/exp format */

    mpegaudio_tableinit();

    for (i = 0; i < 4; i++) {
        if (ff_mpa_quant_bits[i] < 0) {
            for (j = 0; j < (1 << (-ff_mpa_quant_bits[i]+1)); j++) {
                int val1, val2, val3, steps;
                int val = j;
                steps   = ff_mpa_quant_steps[i];
                val1    = val % steps;
                val    /= steps;
                val2    = val % steps;
                val3    = val / steps;
                division_tabs[i][j] = val1 + (val2 << 4) + (val3 << 8);
            }
        }
    }


    for (i = 0; i < 7; i++) {
        float f;
        INTFLOAT v;
        if (i != 6) {
            f = tan((double)i * M_PI / 12.0);
            v = FIXR(f / (1.0 + f));
        } else {
            v = FIXR(1.0);
        }
        is_table[0][    i] = v;
        is_table[1][6 - i] = v;
    }
    /* invalid values */
    for (i = 7; i < 16; i++)
        is_table[0][i] = is_table[1][i] = 0.0;

    for (i = 0; i < 16; i++) {
        double f;
        int e, k;

        for (j = 0; j < 2; j++) {
            e = -(j + 1) * ((i + 1) >> 1);
            f = pow(2.0, e / 4.0);
            k = i & 1;
            is_table_lsf[j][k ^ 1][i] = FIXR(f);
            is_table_lsf[j][k    ][i] = FIXR(1.0);
            ff_dlog(NULL, "is_table_lsf %d %d: %f %f\n",
                    i, j, (float) is_table_lsf[j][0][i],
                    (float) is_table_lsf[j][1][i]);
        }
    }

    for (i = 0; i < 8; i++) {
        float ci, cs, ca;
        ci = ci_table[i];
        cs = 1.0 / sqrt(1.0 + ci * ci);
        ca = cs * ci;
#if !CONFIG_FLOAT
        csa_table[i][0] = FIXHR(cs/4);
        csa_table[i][1] = FIXHR(ca/4);
        csa_table[i][2] = FIXHR(ca/4) + FIXHR(cs/4);
        csa_table[i][3] = FIXHR(ca/4) - FIXHR(cs/4);
#else
        csa_table[i][0] = cs;
        csa_table[i][1] = ca;
        csa_table[i][2] = ca + cs;
        csa_table[i][3] = ca - cs;
#endif
    }
}

static av_cold int decode_init(AVCodecContext * avctx)
{
    static int initialized_tables = 0;
    MPADecodeContext *s = avctx->priv_data;

    if (!initialized_tables) {
        decode_init_static();
        initialized_tables = 1;
    }

    s->avctx = avctx;

    avpriv_float_dsp_init(&s->fdsp, avctx->flags & CODEC_FLAG_BITEXACT);
    ff_mpadsp_init(&s->mpadsp);

    if (avctx->request_sample_fmt == OUT_FMT &&
        avctx->codec_id != AV_CODEC_ID_MP3ON4)
        avctx->sample_fmt = OUT_FMT;
    else
        avctx->sample_fmt = OUT_FMT_P;
    s->err_recognition = avctx->err_recognition;

    if (avctx->codec_id == AV_CODEC_ID_MP3ADU)
        s->adu_mode = 1;

    return 0;
}

#define C3 FIXHR(0.86602540378443864676/2)
#define C4 FIXHR(0.70710678118654752439/2) //0.5 / cos(pi*(9)/36)
#define C5 FIXHR(0.51763809020504152469/2) //0.5 / cos(pi*(5)/36)
#define C6 FIXHR(1.93185165257813657349/4) //0.5 / cos(pi*(15)/36)

/* 12 points IMDCT. We compute it "by hand" by factorizing obvious
   cases. */
static void imdct12(INTFLOAT *out, INTFLOAT *in)
{
    INTFLOAT in0, in1, in2, in3, in4, in5, t1, t2;

    in0  = in[0*3];
    in1  = in[1*3] + in[0*3];
    in2  = in[2*3] + in[1*3];
    in3  = in[3*3] + in[2*3];
    in4  = in[4*3] + in[3*3];
    in5  = in[5*3] + in[4*3];
    in5 += in3;
    in3 += in1;

    in2  = MULH3(in2, C3, 2);
    in3  = MULH3(in3, C3, 4);

    t1   = in0 - in4;
    t2   = MULH3(in1 - in5, C4, 2);

    out[ 7] =
    out[10] = t1 + t2;
    out[ 1] =
    out[ 4] = t1 - t2;

    in0    += SHR(in4, 1);
    in4     = in0 + in2;
    in5    += 2*in1;
    in1     = MULH3(in5 + in3, C5, 1);
    out[ 8] =
    out[ 9] = in4 + in1;
    out[ 2] =
    out[ 3] = in4 - in1;

    in0    -= in2;
    in5     = MULH3(in5 - in3, C6, 2);
    out[ 0] =
    out[ 5] = in0 - in5;
    out[ 6] =
    out[11] = in0 + in5;
}

/* return the number of decoded frames */
static int mp_decode_layer1(MPADecodeContext *s)
{
    int bound, i, v, n, ch, j, mant;
    uint8_t allocation[MPA_MAX_CHANNELS][SBLIMIT];
    uint8_t scale_factors[MPA_MAX_CHANNELS][SBLIMIT];

    if (s->mode == MPA_JSTEREO)
        bound = (s->mode_ext + 1) * 4;
    else
        bound = SBLIMIT;

    /* allocation bits */
    for (i = 0; i < bound; i++) {
        for (ch = 0; ch < s->nb_channels; ch++) {
            allocation[ch][i] = get_bits(&s->gb, 4);
        }
    }
    for (i = bound; i < SBLIMIT; i++)
        allocation[0][i] = get_bits(&s->gb, 4);

    /* scale factors */
    for (i = 0; i < bound; i++) {
        for (ch = 0; ch < s->nb_channels; ch++) {
            if (allocation[ch][i])
                scale_factors[ch][i] = get_bits(&s->gb, 6);
        }
    }
    for (i = bound; i < SBLIMIT; i++) {
        if (allocation[0][i]) {
            scale_factors[0][i] = get_bits(&s->gb, 6);
            scale_factors[1][i] = get_bits(&s->gb, 6);
        }
    }

    /* compute samples */
    for (j = 0; j < 12; j++) {
        for (i = 0; i < bound; i++) {
            for (ch = 0; ch < s->nb_channels; ch++) {
                n = allocation[ch][i];
                if (n) {
                    mant = get_bits(&s->gb, n + 1);
                    v = l1_unscale(n, mant, scale_factors[ch][i]);
                } else {
                    v = 0;
                }
                s->sb_samples[ch][j][i] = v;
            }
        }
        for (i = bound; i < SBLIMIT; i++) {
            n = allocation[0][i];
            if (n) {
                mant = get_bits(&s->gb, n + 1);
                v = l1_unscale(n, mant, scale_factors[0][i]);
                s->sb_samples[0][j][i] = v;
                v = l1_unscale(n, mant, scale_factors[1][i]);
                s->sb_samples[1][j][i] = v;
            } else {
                s->sb_samples[0][j][i] = 0;
                s->sb_samples[1][j][i] = 0;
            }
        }
    }
    return 12;
}

static int mp_decode_layer2(MPADecodeContext *s)
{
    int sblimit; /* number of used subbands */
    const unsigned char *alloc_table;
    int table, bit_alloc_bits, i, j, ch, bound, v;
    unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
    unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
    unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3], *sf;
    int scale, qindex, bits, steps, k, l, m, b;

    /* select decoding table */
    table = ff_mpa_l2_select_table(s->bit_rate / 1000, s->nb_channels,
                                   s->sample_rate, s->lsf);
    sblimit     = ff_mpa_sblimit_table[table];
    alloc_table = ff_mpa_alloc_tables[table];

    if (s->mode == MPA_JSTEREO)
        bound = (s->mode_ext + 1) * 4;
    else
        bound = sblimit;

    ff_dlog(s->avctx, "bound=%d sblimit=%d\n", bound, sblimit);

    /* sanity check */
    if (bound > sblimit)
        bound = sblimit;

    /* parse bit allocation */
    j = 0;
    for (i = 0; i < bound; i++) {
        bit_alloc_bits = alloc_table[j];
        for (ch = 0; ch < s->nb_channels; ch++)
            bit_alloc[ch][i] = get_bits(&s->gb, bit_alloc_bits);
        j += 1 << bit_alloc_bits;
    }
    for (i = bound; i < sblimit; i++) {
        bit_alloc_bits = alloc_table[j];
        v = get_bits(&s->gb, bit_alloc_bits);
        bit_alloc[0][i] = v;
        bit_alloc[1][i] = v;
        j += 1 << bit_alloc_bits;
    }

    /* scale codes */
    for (i = 0; i < sblimit; i++) {
        for (ch = 0; ch < s->nb_channels; ch++) {
            if (bit_alloc[ch][i])
                scale_code[ch][i] = get_bits(&s->gb, 2);
        }
    }

    /* scale factors */
    for (i = 0; i < sblimit; i++) {
        for (ch = 0; ch < s->nb_channels; ch++) {
            if (bit_alloc[ch][i]) {
                sf = scale_factors[ch][i];
                switch (scale_code[ch][i]) {
                default:
                case 0:
                    sf[0] = get_bits(&s->gb, 6);
                    sf[1] = get_bits(&s->gb, 6);
                    sf[2] = get_bits(&s->gb, 6);
                    break;
                case 2:
                    sf[0] = get_bits(&s->gb, 6);
                    sf[1] = sf[0];
                    sf[2] = sf[0];
                    break;
                case 1:
                    sf[0] = get_bits(&s->gb, 6);
                    sf[2] = get_bits(&s->gb, 6);
                    sf[1] = sf[0];
                    break;
                case 3:
                    sf[0] = get_bits(&s->gb, 6);
                    sf[2] = get_bits(&s->gb, 6);
                    sf[1] = sf[2];
                    break;
                }
            }
        }
    }

    /* samples */
    for (k = 0; k < 3; k++) {
        for (l = 0; l < 12; l += 3) {
            j = 0;
            for (i = 0; i < bound; i++) {
                bit_alloc_bits = alloc_table[j];
                for (ch = 0; ch < s->nb_channels; ch++) {
                    b = bit_alloc[ch][i];
                    if (b) {
                        scale = scale_factors[ch][i][k];
                        qindex = alloc_table[j+b];
                        bits = ff_mpa_quant_bits[qindex];
                        if (bits < 0) {
                            int v2;
                            /* 3 values at the same time */
                            v = get_bits(&s->gb, -bits);
                            v2 = division_tabs[qindex][v];
                            steps  = ff_mpa_quant_steps[qindex];

                            s->sb_samples[ch][k * 12 + l + 0][i] =
                                l2_unscale_group(steps,  v2       & 15, scale);
                            s->sb_samples[ch][k * 12 + l + 1][i] =
                                l2_unscale_group(steps, (v2 >> 4) & 15, scale);
                            s->sb_samples[ch][k * 12 + l + 2][i] =
                                l2_unscale_group(steps,  v2 >> 8      , scale);
                        } else {
                            for (m = 0; m < 3; m++) {
                                v = get_bits(&s->gb, bits);
                                v = l1_unscale(bits - 1, v, scale);
                                s->sb_samples[ch][k * 12 + l + m][i] = v;
                            }
                        }
                    } else {
                        s->sb_samples[ch][k * 12 + l + 0][i] = 0;
                        s->sb_samples[ch][k * 12 + l + 1][i] = 0;
                        s->sb_samples[ch][k * 12 + l + 2][i] = 0;
                    }
                }
                /* next subband in alloc table */
                j += 1 << bit_alloc_bits;
            }
            /* XXX: find a way to avoid this duplication of code */
            for (i = bound; i < sblimit; i++) {
                bit_alloc_bits = alloc_table[j];
                b = bit_alloc[0][i];
                if (b) {
                    int mant, scale0, scale1;
                    scale0 = scale_factors[0][i][k];
                    scale1 = scale_factors[1][i][k];
                    qindex = alloc_table[j+b];
                    bits = ff_mpa_quant_bits[qindex];
                    if (bits < 0) {
                        /* 3 values at the same time */
                        v = get_bits(&s->gb, -bits);
                        steps = ff_mpa_quant_steps[qindex];
                        mant = v % steps;
                        v = v / steps;
                        s->sb_samples[0][k * 12 + l + 0][i] =
                            l2_unscale_group(steps, mant, scale0);
                        s->sb_samples[1][k * 12 + l + 0][i] =
                            l2_unscale_group(steps, mant, scale1);
                        mant = v % steps;
                        v = v / steps;
                        s->sb_samples[0][k * 12 + l + 1][i] =
                            l2_unscale_group(steps, mant, scale0);
                        s->sb_samples[1][k * 12 + l + 1][i] =
                            l2_unscale_group(steps, mant, scale1);
                        s->sb_samples[0][k * 12 + l + 2][i] =
                            l2_unscale_group(steps, v, scale0);
                        s->sb_samples[1][k * 12 + l + 2][i] =
                            l2_unscale_group(steps, v, scale1);
                    } else {
                        for (m = 0; m < 3; m++) {
                            mant = get_bits(&s->gb, bits);
                            s->sb_samples[0][k * 12 + l + m][i] =
                                l1_unscale(bits - 1, mant, scale0);
                            s->sb_samples[1][k * 12 + l + m][i] =
                                l1_unscale(bits - 1, mant, scale1);
                        }
                    }
                } else {
                    s->sb_samples[0][k * 12 + l + 0][i] = 0;
                    s->sb_samples[0][k * 12 + l + 1][i] = 0;
                    s->sb_samples[0][k * 12 + l + 2][i] = 0;
                    s->sb_samples[1][k * 12 + l + 0][i] = 0;
                    s->sb_samples[1][k * 12 + l + 1][i] = 0;
                    s->sb_samples[1][k * 12 + l + 2][i] = 0;
                }
                /* next subband in alloc table */
                j += 1 << bit_alloc_bits;
            }
            /* fill remaining samples to zero */
            for (i = sblimit; i < SBLIMIT; i++) {
                for (ch = 0; ch < s->nb_channels; ch++) {
                    s->sb_samples[ch][k * 12 + l + 0][i] = 0;
                    s->sb_samples[ch][k * 12 + l + 1][i] = 0;
                    s->sb_samples[ch][k * 12 + l + 2][i] = 0;
                }
            }
        }
    }
    return 3 * 12;
}

#define SPLIT(dst,sf,n)             \
    if (n == 3) {                   \
        int m = (sf * 171) >> 9;    \
        dst   = sf - 3 * m;         \
        sf    = m;                  \
    } else if (n == 4) {            \
        dst  = sf & 3;              \
        sf >>= 2;                   \
    } else if (n == 5) {            \
        int m = (sf * 205) >> 10;   \
        dst   = sf - 5 * m;         \
        sf    = m;                  \
    } else if (n == 6) {            \
        int m = (sf * 171) >> 10;   \
        dst   = sf - 6 * m;         \
        sf    = m;                  \
    } else {                        \
        dst = 0;                    \
    }

static av_always_inline void lsf_sf_expand(int *slen, int sf, int n1, int n2,
                                           int n3)
{
    SPLIT(slen[3], sf, n3)
    SPLIT(slen[2], sf, n2)
    SPLIT(slen[1], sf, n1)
    slen[0] = sf;
}

static void exponents_from_scale_factors(MPADecodeContext *s, GranuleDef *g,
                                         int16_t *exponents)
{
    const uint8_t *bstab, *pretab;
    int len, i, j, k, l, v0, shift, gain, gains[3];
    int16_t *exp_ptr;

    exp_ptr = exponents;
    gain    = g->global_gain - 210;
    shift   = g->scalefac_scale + 1;

    bstab  = band_size_long[s->sample_rate_index];
    pretab = mpa_pretab[g->preflag];
    for (i = 0; i < g->long_end; i++) {
        v0 = gain - ((g->scale_factors[i] + pretab[i]) << shift) + 400;
        len = bstab[i];
        for (j = len; j > 0; j--)
            *exp_ptr++ = v0;
    }

    if (g->short_start < 13) {
        bstab    = band_size_short[s->sample_rate_index];
        gains[0] = gain - (g->subblock_gain[0] << 3);
        gains[1] = gain - (g->subblock_gain[1] << 3);
        gains[2] = gain - (g->subblock_gain[2] << 3);
        k        = g->long_end;
        for (i = g->short_start; i < 13; i++) {
            len = bstab[i];
            for (l = 0; l < 3; l++) {
                v0 = gains[l] - (g->scale_factors[k++] << shift) + 400;
                for (j = len; j > 0; j--)
                    *exp_ptr++ = v0;
            }
        }
    }
}

/* handle n = 0 too */
static inline int get_bitsz(GetBitContext *s, int n)
{
    return n ? get_bits(s, n) : 0;
}


static void switch_buffer(MPADecodeContext *s, int *pos, int *end_pos,
                          int *end_pos2)
{
    if (s->in_gb.buffer && *pos >= s->gb.size_in_bits) {
        s->gb           = s->in_gb;
        s->in_gb.buffer = NULL;
        assert((get_bits_count(&s->gb) & 7) == 0);
        skip_bits_long(&s->gb, *pos - *end_pos);
        *end_pos2 =
        *end_pos  = *end_pos2 + get_bits_count(&s->gb) - *pos;
        *pos      = get_bits_count(&s->gb);
    }
}

/* Following is a optimized code for
            INTFLOAT v = *src
            if(get_bits1(&s->gb))
                v = -v;
            *dst = v;
*/
#if CONFIG_FLOAT
#define READ_FLIP_SIGN(dst,src)                     \
    v = AV_RN32A(src) ^ (get_bits1(&s->gb) << 31);  \
    AV_WN32A(dst, v);
#else
#define READ_FLIP_SIGN(dst,src)     \
    v      = -get_bits1(&s->gb);    \
    *(dst) = (*(src) ^ v) - v;
#endif

static int huffman_decode(MPADecodeContext *s, GranuleDef *g,
                          int16_t *exponents, int end_pos2)
{
    int s_index;
    int i;
    int last_pos, bits_left;
    VLC *vlc;
    int end_pos = FFMIN(end_pos2, s->gb.size_in_bits);

    /* low frequencies (called big values) */
    s_index = 0;
    for (i = 0; i < 3; i++) {
        int j, k, l, linbits;
        j = g->region_size[i];
        if (j == 0)
            continue;
        /* select vlc table */
        k       = g->table_select[i];
        l       = mpa_huff_data[k][0];
        linbits = mpa_huff_data[k][1];
        vlc     = &huff_vlc[l];

        if (!l) {
            memset(&g->sb_hybrid[s_index], 0, sizeof(*g->sb_hybrid) * 2 * j);
            s_index += 2 * j;
            continue;
        }

        /* read huffcode and compute each couple */
        for (; j > 0; j--) {
            int exponent, x, y;
            int v;
            int pos = get_bits_count(&s->gb);

            if (pos >= end_pos){
                switch_buffer(s, &pos, &end_pos, &end_pos2);
                if (pos >= end_pos)
                    break;
            }
            y = get_vlc2(&s->gb, vlc->table, 7, 3);

            if (!y) {
                g->sb_hybrid[s_index  ] =
                g->sb_hybrid[s_index+1] = 0;
                s_index += 2;
                continue;
            }

            exponent= exponents[s_index];

            ff_dlog(s->avctx, "region=%d n=%d x=%d y=%d exp=%d\n",
                    i, g->region_size[i] - j, x, y, exponent);
            if (y & 16) {
                x = y >> 5;
                y = y & 0x0f;
                if (x < 15) {
                    READ_FLIP_SIGN(g->sb_hybrid + s_index, RENAME(expval_table)[exponent] + x)
                } else {
                    x += get_bitsz(&s->gb, linbits);
                    v  = l3_unscale(x, exponent);
                    if (get_bits1(&s->gb))
                        v = -v;
                    g->sb_hybrid[s_index] = v;
                }
                if (y < 15) {
                    READ_FLIP_SIGN(g->sb_hybrid + s_index + 1, RENAME(expval_table)[exponent] + y)
                } else {
                    y += get_bitsz(&s->gb, linbits);
                    v  = l3_unscale(y, exponent);
                    if (get_bits1(&s->gb))
                        v = -v;
                    g->sb_hybrid[s_index+1] = v;
                }
            } else {
                x = y >> 5;
                y = y & 0x0f;
                x += y;
                if (x < 15) {
                    READ_FLIP_SIGN(g->sb_hybrid + s_index + !!y, RENAME(expval_table)[exponent] + x)
                } else {
                    x += get_bitsz(&s->gb, linbits);
                    v  = l3_unscale(x, exponent);
                    if (get_bits1(&s->gb))
                        v = -v;
                    g->sb_hybrid[s_index+!!y] = v;
                }
                g->sb_hybrid[s_index + !y] = 0;
            }
            s_index += 2;
        }
    }

    /* high frequencies */
    vlc = &huff_quad_vlc[g->count1table_select];
    last_pos = 0;
    while (s_index <= 572) {
        int pos, code;
        pos = get_bits_count(&s->gb);
        if (pos >= end_pos) {
            if (pos > end_pos2 && last_pos) {
                /* some encoders generate an incorrect size for this
                   part. We must go back into the data */
                s_index -= 4;
                skip_bits_long(&s->gb, last_pos - pos);
                av_log(s->avctx, AV_LOG_INFO, "overread, skip %d enddists: %d %d\n", last_pos - pos, end_pos-pos, end_pos2-pos);
                if(s->err_recognition & AV_EF_BITSTREAM)
                    s_index=0;
                break;
            }
            switch_buffer(s, &pos, &end_pos, &end_pos2);
            if (pos >= end_pos)
                break;
        }
        last_pos = pos;

        code = get_vlc2(&s->gb, vlc->table, vlc->bits, 1);
        ff_dlog(s->avctx, "t=%d code=%d\n", g->count1table_select, code);
        g->sb_hybrid[s_index+0] =
        g->sb_hybrid[s_index+1] =
        g->sb_hybrid[s_index+2] =
        g->sb_hybrid[s_index+3] = 0;
        while (code) {
            static const int idxtab[16] = { 3,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0 };
            int v;
            int pos = s_index + idxtab[code];
            code   ^= 8 >> idxtab[code];
            READ_FLIP_SIGN(g->sb_hybrid + pos, RENAME(exp_table)+exponents[pos])
        }
        s_index += 4;
    }
    /* skip extension bits */
    bits_left = end_pos2 - get_bits_count(&s->gb);
    if (bits_left < 0 && (s->err_recognition & AV_EF_BUFFER)) {
        av_log(s->avctx, AV_LOG_ERROR, "bits_left=%d\n", bits_left);
        s_index=0;
    } else if (bits_left > 0 && (s->err_recognition & AV_EF_BUFFER)) {
        av_log(s->avctx, AV_LOG_ERROR, "bits_left=%d\n", bits_left);
        s_index = 0;
    }
    memset(&g->sb_hybrid[s_index], 0, sizeof(*g->sb_hybrid) * (576 - s_index));
    skip_bits_long(&s->gb, bits_left);

    i = get_bits_count(&s->gb);
    switch_buffer(s, &i, &end_pos, &end_pos2);

    return 0;
}

/* Reorder short blocks from bitstream order to interleaved order. It
   would be faster to do it in parsing, but the code would be far more
   complicated */
static void reorder_block(MPADecodeContext *s, GranuleDef *g)
{
    int i, j, len;
    INTFLOAT *ptr, *dst, *ptr1;
    INTFLOAT tmp[576];

    if (g->block_type != 2)
        return;

    if (g->switch_point) {
        if (s->sample_rate_index != 8)
            ptr = g->sb_hybrid + 36;
        else
            ptr = g->sb_hybrid + 72;
    } else {