annotate mpegaudiodec.c @ 11939:ef338bd70180 libavcodec

SSE-optimized MP3 floating point windowing functions

/*
 * MPEG Audio decoder
 * Copyright (c) 2001, 2002 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
 * MPEG Audio decoder.
 */

#include "avcodec.h"
#include "get_bits.h"
#include "dsputil.h"

/*
 * TODO:
 *  - in low precision mode, use more 16 bit multiplies in synth filter
 *  - test lsf / mpeg25 extensively.
 */

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

#include "mathops.h"

#if CONFIG_FLOAT
#   define SHR(a,b)       ((a)*(1.0f/(1<<(b))))
#   define compute_antialias compute_antialias_float
#   define FIXR_OLD(a)    ((int)((a) * FRAC_ONE + 0.5))
#   define FIXR(x)        ((float)(x))
#   define FIXHR(x)       ((float)(x))
#   define MULH3(x, y, s) ((s)*(y)*(x))
#   define MULLx(x, y, s) ((y)*(x))
#   define RENAME(a) a ## _float
#else
#   define SHR(a,b)       ((a)>>(b))
#   define compute_antialias compute_antialias_integer
/* WARNING: only correct for posititive numbers */
#   define FIXR_OLD(a)    ((int)((a) * FRAC_ONE + 0.5))
#   define FIXR(a)        ((int)((a) * FRAC_ONE + 0.5))
#   define FIXHR(a)       ((int)((a) * (1LL<<32) + 0.5))
#   define MULH3(x, y, s) MULH((s)*(x), y)
#   define MULLx(x, y, s) MULL(x,y,s)
#   define RENAME(a)      a
#endif

/****************/

#define HEADER_SIZE 4

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

static void compute_antialias_integer(MPADecodeContext *s, GranuleDef *g);
static void compute_antialias_float(MPADecodeContext *s, GranuleDef *g);
static void apply_window_mp3_c(MPA_INT *synth_buf, MPA_INT *window,
                               int *dither_state, OUT_INT *samples, int incr);

/* 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 int32_t csa_table[8][4];
static float csa_table_float[8][4];
static INTFLOAT mdct_win[8][36];

/* 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 */
};

DECLARE_ALIGNED(16, MPA_INT, RENAME(ff_mpa_synth_window))[512+256];

/**
 * Convert region offsets to region sizes and truncate
 * size to big_values.
 */
static void ff_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 ff_init_short_region(MPADecodeContext *s, GranuleDef *g){
    if (g->block_type == 2)
        g->region_size[0] = (36 / 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 ff_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 ff_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 48 first
                exponents as long blocks (XXX: check this!) */
            if (s->sample_rate_index <= 2)
                g->long_end = 8;
            else if (s->sample_rate_index != 8)
                g->long_end = 6;
            else
                g->long_end = 4; /* 8000 Hz */

            g->short_start = 2 + (s->sample_rate_index != 8);
        } 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;
}

/* all integer n^(4/3) computation code */
#define DEV_ORDER 13

#define POW_FRAC_BITS 24
#define POW_FRAC_ONE    (1 << POW_FRAC_BITS)
#define POW_FIX(a)   ((int)((a) * POW_FRAC_ONE))
#define POW_MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> POW_FRAC_BITS)

static int dev_4_3_coefs[DEV_ORDER];

#if 0 /* unused */
static int pow_mult3[3] = {
    POW_FIX(1.0),
    POW_FIX(1.25992104989487316476),
    POW_FIX(1.58740105196819947474),
};
#endif

static av_cold void int_pow_init(void)
{
    int i, a;

    a = POW_FIX(1.0);
    for(i=0;i<DEV_ORDER;i++) {
        a = POW_MULL(a, POW_FIX(4.0 / 3.0) - i * POW_FIX(1.0)) / (i + 1);
        dev_4_3_coefs[i] = a;
    }
}

#if 0 /* unused, remove? */
/* return the mantissa and the binary exponent */
static int int_pow(int i, int *exp_ptr)
{
    int e, er, eq, j;
    int a, a1;

    /* renormalize */
    a = i;
    e = POW_FRAC_BITS;
    while (a < (1 << (POW_FRAC_BITS - 1))) {
        a = a << 1;
        e--;
    }
    a -= (1 << POW_FRAC_BITS);
    a1 = 0;
    for(j = DEV_ORDER - 1; j >= 0; j--)
        a1 = POW_MULL(a, dev_4_3_coefs[j] + a1);
    a = (1 << POW_FRAC_BITS) + a1;
    /* exponent compute (exact) */
    e = e * 4;
    er = e % 3;
    eq = e / 3;
    a = POW_MULL(a, pow_mult3[er]);
    while (a >= 2 * POW_FRAC_ONE) {
        a = a >> 1;
        eq++;
    }
    /* convert to float */
    while (a < POW_FRAC_ONE) {
        a = a << 1;
        eq--;
    }
    /* now POW_FRAC_ONE <= a < 2 * POW_FRAC_ONE */
#if POW_FRAC_BITS > FRAC_BITS
    a = (a + (1 << (POW_FRAC_BITS - FRAC_BITS - 1))) >> (POW_FRAC_BITS - FRAC_BITS);
    /* correct overflow */
    if (a >= 2 * (1 << FRAC_BITS)) {
        a = a >> 1;
        eq++;
    }
#endif
    *exp_ptr = eq;
    return a;
}
#endif

static av_cold int decode_init(AVCodecContext * avctx)
{
    MPADecodeContext *s = avctx->priv_data;
    static int init=0;
    int i, j, k;

    s->avctx = avctx;
    s->apply_window_mp3 = apply_window_mp3_c;
#if HAVE_MMX
    ff_mpegaudiodec_init_mmx(s);
#endif
    avctx->sample_fmt= OUT_FMT;
    s->error_recognition= avctx->error_recognition;

    if (!init && !avctx->parse_only) {
        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);
            dprintf(avctx, "%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];
            uint16_t tmp_codes[512];

            memset(tmp_bits , 0, sizeof(tmp_bits ));
            memset(tmp_codes, 0, sizeof(tmp_codes));

            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 */

        int_pow_init();
        mpegaudio_tableinit();

        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);
                dprintf(avctx, "is_table_lsf %d %d: %x %x\n",
                        i, j, is_table_lsf[j][0][i], 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;
            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);
            csa_table_float[i][0] = cs;
            csa_table_float[i][1] = ca;
            csa_table_float[i][2] = ca + cs;
            csa_table_float[i][3] = ca - cs;
        }

        /* compute mdct windows */
        for(i=0;i<36;i++) {
            for(j=0; j<4; j++){
                double d;

                if(j==2 && i%3 != 1)
                    continue;

                d= sin(M_PI * (i + 0.5) / 36.0);
                if(j==1){
                    if     (i>=30) d= 0;
                    else if(i>=24) d= sin(M_PI * (i - 18 + 0.5) / 12.0);
                    else if(i>=18) d= 1;
                }else if(j==3){
                    if     (i<  6) d= 0;
                    else if(i< 12) d= sin(M_PI * (i -  6 + 0.5) / 12.0);
                    else if(i< 18) d= 1;
                }
                //merge last stage of imdct into the window coefficients
                d*= 0.5 / cos(M_PI*(2*i + 19)/72);

                if(j==2)
                    mdct_win[j][i/3] = FIXHR((d / (1<<5)));
                else
                    mdct_win[j][i  ] = FIXHR((d / (1<<5)));
            }
        }

        /* NOTE: we do frequency inversion adter the MDCT by changing
           the sign of the right window coefs */
        for(j=0;j<4;j++) {
            for(i=0;i<36;i+=2) {
                mdct_win[j + 4][i] = mdct_win[j][i];
                mdct_win[j + 4][i + 1] = -mdct_win[j][i + 1];
            }
        }

        init = 1;
    }

    if (avctx->codec_id == CODEC_ID_MP3ADU)
        s->adu_mode = 1;
    return 0;
}

/* tab[i][j] = 1.0 / (2.0 * cos(pi*(2*k+1) / 2^(6 - j))) */

/* cos(i*pi/64) */

#define COS0_0  FIXHR(0.50060299823519630134/2)
#define COS0_1  FIXHR(0.50547095989754365998/2)
#define COS0_2  FIXHR(0.51544730992262454697/2)
#define COS0_3  FIXHR(0.53104259108978417447/2)
#define COS0_4  FIXHR(0.55310389603444452782/2)
#define COS0_5  FIXHR(0.58293496820613387367/2)
#define COS0_6  FIXHR(0.62250412303566481615/2)
#define COS0_7  FIXHR(0.67480834145500574602/2)
#define COS0_8  FIXHR(0.74453627100229844977/2)
#define COS0_9  FIXHR(0.83934964541552703873/2)
#define COS0_10 FIXHR(0.97256823786196069369/2)
#define COS0_11 FIXHR(1.16943993343288495515/4)
#define COS0_12 FIXHR(1.48416461631416627724/4)
#define COS0_13 FIXHR(2.05778100995341155085/8)
#define COS0_14 FIXHR(3.40760841846871878570/8)
#define COS0_15 FIXHR(10.19000812354805681150/32)

#define COS1_0 FIXHR(0.50241928618815570551/2)
#define COS1_1 FIXHR(0.52249861493968888062/2)
#define COS1_2 FIXHR(0.56694403481635770368/2)
#define COS1_3 FIXHR(0.64682178335999012954/2)
#define COS1_4 FIXHR(0.78815462345125022473/2)
#define COS1_5 FIXHR(1.06067768599034747134/4)
#define COS1_6 FIXHR(1.72244709823833392782/4)
#define COS1_7 FIXHR(5.10114861868916385802/16)

#define COS2_0 FIXHR(0.50979557910415916894/2)
#define COS2_1 FIXHR(0.60134488693504528054/2)
#define COS2_2 FIXHR(0.89997622313641570463/2)
#define COS2_3 FIXHR(2.56291544774150617881/8)

#define COS3_0 FIXHR(0.54119610014619698439/2)
#define COS3_1 FIXHR(1.30656296487637652785/4)

#define COS4_0 FIXHR(0.70710678118654752439/2)

/* butterfly operator */
#define BF(a, b, c, s)\
{\
    tmp0 = val##a + val##b;\
    tmp1 = val##a - val##b;\
    val##a = tmp0;\
    val##b = MULH3(tmp1, c, 1<<(s));\
}

#define BF0(a, b, c, s)\
{\
    tmp0 = tab[a] + tab[b];\
    tmp1 = tab[a] - tab[b];\
    val##a = tmp0;\
    val##b = MULH3(tmp1, c, 1<<(s));\
}

#define BF1(a, b, c, d)\
{\
    BF(a, b, COS4_0, 1);\
    BF(c, d,-COS4_0, 1);\
    val##c += val##d;\
}

#define BF2(a, b, c, d)\
{\
    BF(a, b, COS4_0, 1);\
    BF(c, d,-COS4_0, 1);\
    val##c += val##d;\
    val##a += val##c;\
    val##c += val##b;\
    val##b += val##d;\
}

#define ADD(a, b) val##a += val##b

/* DCT32 without 1/sqrt(2) coef zero scaling. */
static void dct32(INTFLOAT *out, const INTFLOAT *tab)
{
    INTFLOAT tmp0, tmp1;

    INTFLOAT val0 , val1 , val2 , val3 , val4 , val5 , val6 , val7 ,
             val8 , val9 , val10, val11, val12, val13, val14, val15,
             val16, val17, val18, val19, val20, val21, val22, val23,
             val24, val25, val26, val27, val28, val29, val30, val31;

    /* pass 1 */
    BF0( 0, 31, COS0_0 , 1);
    BF0(15, 16, COS0_15, 5);
    /* pass 2 */
    BF( 0, 15, COS1_0 , 1);
    BF(16, 31,-COS1_0 , 1);
    /* pass 1 */
    BF0( 7, 24, COS0_7 , 1);
    BF0( 8, 23, COS0_8 , 1);
    /* pass 2 */
    BF( 7,  8, COS1_7 , 4);
    BF(23, 24,-COS1_7 , 4);
    /* pass 3 */
    BF( 0,  7, COS2_0 , 1);
    BF( 8, 15,-COS2_0 , 1);
    BF(16, 23, COS2_0 , 1);
    BF(24, 31,-COS2_0 , 1);
    /* pass 1 */
    BF0( 3, 28, COS0_3 , 1);
    BF0(12, 19, COS0_12, 2);
    /* pass 2 */
    BF( 3, 12, COS1_3 , 1);
    BF(19, 28,-COS1_3 , 1);
    /* pass 1 */
    BF0( 4, 27, COS0_4 , 1);
    BF0(11, 20, COS0_11, 2);
    /* pass 2 */
    BF( 4, 11, COS1_4 , 1);
    BF(20, 27,-COS1_4 , 1);
    /* pass 3 */
    BF( 3,  4, COS2_3 , 3);
    BF(11, 12,-COS2_3 , 3);
    BF(19, 20, COS2_3 , 3);
    BF(27, 28,-COS2_3 , 3);
    /* pass 4 */
    BF( 0,  3, COS3_0 , 1);
    BF( 4,  7,-COS3_0 , 1);
    BF( 8, 11, COS3_0 , 1);
    BF(12, 15,-COS3_0 , 1);
    BF(16, 19, COS3_0 , 1);
    BF(20, 23,-COS3_0 , 1);
    BF(24, 27, COS3_0 , 1);
    BF(28, 31,-COS3_0 , 1);



    /* pass 1 */
    BF0( 1, 30, COS0_1 , 1);
    BF0(14, 17, COS0_14, 3);
    /* pass 2 */
    BF( 1, 14, COS1_1 , 1);
    BF(17, 30,-COS1_1 , 1);
    /* pass 1 */
    BF0( 6, 25, COS0_6 , 1);
    BF0( 9, 22, COS0_9 , 1);
    /* pass 2 */
    BF( 6,  9, COS1_6 , 2);
    BF(22, 25,-COS1_6 , 2);
    /* pass 3 */
    BF( 1,  6, COS2_1 , 1);
    BF( 9, 14,-COS2_1 , 1);
    BF(17, 22, COS2_1 , 1);
    BF(25, 30,-COS2_1 , 1);

    /* pass 1 */
    BF0( 2, 29, COS0_2 , 1);
    BF0(13, 18, COS0_13, 3);
    /* pass 2 */
    BF( 2, 13, COS1_2 , 1);
    BF(18, 29,-COS1_2 , 1);
    /* pass 1 */
    BF0( 5, 26, COS0_5 , 1);
    BF0(10, 21, COS0_10, 1);
    /* pass 2 */
    BF( 5, 10, COS1_5 , 2);
    BF(21, 26,-COS1_5 , 2);
    /* pass 3 */
    BF( 2,  5, COS2_2 , 1);
    BF(10, 13,-COS2_2 , 1);
    BF(18, 21, COS2_2 , 1);
    BF(26, 29,-COS2_2 , 1);
    /* pass 4 */
    BF( 1,  2, COS3_1 , 2);
    BF( 5,  6,-COS3_1 , 2);
    BF( 9, 10, COS3_1 , 2);
    BF(13, 14,-COS3_1 , 2);
    BF(17, 18, COS3_1 , 2);
    BF(21, 22,-COS3_1 , 2);
    BF(25, 26, COS3_1 , 2);
    BF(29, 30,-COS3_1 , 2);

    /* pass 5 */
    BF1( 0,  1,  2,  3);
    BF2( 4,  5,  6,  7);
    BF1( 8,  9, 10, 11);
    BF2(12, 13, 14, 15);
    BF1(16, 17, 18, 19);
    BF2(20, 21, 22, 23);
    BF1(24, 25, 26, 27);
    BF2(28, 29, 30, 31);

    /* pass 6 */

    ADD( 8, 12);
    ADD(12, 10);
    ADD(10, 14);
    ADD(14,  9);
    ADD( 9, 13);
    ADD(13, 11);
    ADD(11, 15);

    out[ 0] = val0;
    out[16] = val1;
    out[ 8] = val2;
    out[24] = val3;
    out[ 4] = val4;
    out[20] = val5;
    out[12] = val6;
    out[28] = val7;
    out[ 2] = val8;
    out[18] = val9;
    out[10] = val10;
    out[26] = val11;
    out[ 6] = val12;
    out[22] = val13;
    out[14] = val14;
    out[30] = val15;

    ADD(24, 28);
    ADD(28, 26);
    ADD(26, 30);
    ADD(30, 25);
    ADD(25, 29);
    ADD(29, 27);
    ADD(27, 31);

    out[ 1] = val16 + val24;
    out[17] = val17 + val25;
    out[ 9] = val18 + val26;
    out[25] = val19 + val27;
    out[ 5] = val20 + val28;
    out[21] = val21 + val29;
    out[13] = val22 + val30;
    out[29] = val23 + val31;
    out[ 3] = val24 + val20;
    out[19] = val25 + val21;
    out[11] = val26 + val22;
    out[27] = val27 + val23;
    out[ 7] = val28 + val18;
    out[23] = val29 + val19;
    out[15] = val30 + val17;
    out[31] = val31;
}

#if CONFIG_FLOAT
static inline float round_sample(float *sum)
{
    float sum1=*sum;
    *sum = 0;
    return sum1;
}

/* signed 16x16 -> 32 multiply add accumulate */
#define MACS(rt, ra, rb) rt+=(ra)*(rb)

/* signed 16x16 -> 32 multiply */
#define MULS(ra, rb) ((ra)*(rb))

#define MLSS(rt, ra, rb) rt-=(ra)*(rb)

#elif FRAC_BITS <= 15

static inline int round_sample(int *sum)
{
    int sum1;
    sum1 = (*sum) >> OUT_SHIFT;
    *sum &= (1<<OUT_SHIFT)-1;
    return av_clip(sum1, OUT_MIN, OUT_MAX);
}

/* signed 16x16 -> 32 multiply add accumulate */
#define MACS(rt, ra, rb) MAC16(rt, ra, rb)

/* signed 16x16 -> 32 multiply */
#define MULS(ra, rb) MUL16(ra, rb)

#define MLSS(rt, ra, rb) MLS16(rt, ra, rb)

#else

static inline int round_sample(int64_t *sum)
{
    int sum1;
    sum1 = (int)((*sum) >> OUT_SHIFT);
    *sum &= (1<<OUT_SHIFT)-1;
    return av_clip(sum1, OUT_MIN, OUT_MAX);
}

#   define MULS(ra, rb) MUL64(ra, rb)
#   define MACS(rt, ra, rb) MAC64(rt, ra, rb)
#   define MLSS(rt, ra, rb) MLS64(rt, ra, rb)
#endif

#define SUM8(op, sum, w, p)               \
{                                         \
    op(sum, (w)[0 * 64], (p)[0 * 64]);    \
    op(sum, (w)[1 * 64], (p)[1 * 64]);    \
    op(sum, (w)[2 * 64], (p)[2 * 64]);    \
    op(sum, (w)[3 * 64], (p)[3 * 64]);    \
    op(sum, (w)[4 * 64], (p)[4 * 64]);    \
    op(sum, (w)[5 * 64], (p)[5 * 64]);    \
    op(sum, (w)[6 * 64], (p)[6 * 64]);    \
    op(sum, (w)[7 * 64], (p)[7 * 64]);    \
}

#define SUM8P2(sum1, op1, sum2, op2, w1, w2, p) \
{                                               \
    INTFLOAT tmp;\
    tmp = p[0 * 64];\
    op1(sum1, (w1)[0 * 64], tmp);\
    op2(sum2, (w2)[0 * 64], tmp);\
    tmp = p[1 * 64];\
    op1(sum1, (w1)[1 * 64], tmp);\
    op2(sum2, (w2)[1 * 64], tmp);\
    tmp = p[2 * 64];\
    op1(sum1, (w1)[2 * 64], tmp);\
    op2(sum2, (w2)[2 * 64], tmp);\
    tmp = p[3 * 64];\
    op1(sum1, (w1)[3 * 64], tmp);\
    op2(sum2, (w2)[3 * 64], tmp);\
    tmp = p[4 * 64];\
    op1(sum1, (w1)[4 * 64], tmp);\
    op2(sum2, (w2)[4 * 64], tmp);\
    tmp = p[5 * 64];\
    op1(sum1, (w1)[5 * 64], tmp);\
    op2(sum2, (w2)[5 * 64], tmp);\
    tmp = p[6 * 64];\
    op1(sum1, (w1)[6 * 64], tmp);\
    op2(sum2, (w2)[6 * 64], tmp);\
    tmp = p[7 * 64];\
    op1(sum1, (w1)[7 * 64], tmp);\
    op2(sum2, (w2)[7 * 64], tmp);\
}

void av_cold RENAME(ff_mpa_synth_init)(MPA_INT *window)
{
    int i, j;

    /* max = 18760, max sum over all 16 coefs : 44736 */
    for(i=0;i<257;i++) {
        INTFLOAT v;
        v = ff_mpa_enwindow[i];
#if CONFIG_FLOAT
        v *= 1.0 / (1LL<<(16 + FRAC_BITS));
#elif WFRAC_BITS < 16
        v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
#endif
        window[i] = v;
        if ((i & 63) != 0)
            v = -v;
        if (i != 0)
            window[512 - i] = v;
    }

    // Needed for avoiding shuffles in ASM implementations
    for(i=0; i < 8; i++)
        for(j=0; j < 16; j++)
            window[512+16*i+j] = window[64*i+32-j];

    for(i=0; i < 8; i++)
        for(j=0; j < 16; j++)
            window[512+128+16*i+j] = window[64*i+48-j];
}

static void apply_window_mp3_c(MPA_INT *synth_buf, MPA_INT *window,
                               int *dither_state, OUT_INT *samples, int incr)
{
    register const MPA_INT *w, *w2, *p;
    int j;
    OUT_INT *samples2;
#if CONFIG_FLOAT
    float sum, sum2;
#elif FRAC_BITS <= 15
    int sum, sum2;
#else
    int64_t sum, sum2;
#endif

    /* copy to avoid wrap */
    memcpy(synth_buf + 512, synth_buf, 32 * sizeof(*synth_buf));

    samples2 = samples + 31 * incr;
    w = window;
    w2 = window + 31;

    sum = *dither_state;
    p = synth_buf + 16;
    SUM8(MACS, sum, w, p);
    p = synth_buf + 48;
    SUM8(MLSS, sum, w + 32, p);
    *samples = round_sample(&sum);
    samples += incr;
    w++;

    /* we calculate two samples at the same time to avoid one memory
       access per two sample */
    for(j=1;j<16;j++) {
        sum2 = 0;
        p = synth_buf + 16 + j;
        SUM8P2(sum, MACS, sum2, MLSS, w, w2, p);
        p = synth_buf + 48 - j;
        SUM8P2(sum, MLSS, sum2, MLSS, w + 32, w2 + 32, p);

        *samples = round_sample(&sum);
        samples += incr;
        sum += sum2;
        *samples2 = round_sample(&sum);
        samples2 -= incr;
        w++;
        w2--;
    }

    p = synth_buf + 32;
    SUM8(MLSS, sum, w + 32, p);
    *samples = round_sample(&sum);
    *dither_state= sum;
}


/* 32 sub band synthesis filter. Input: 32 sub band samples, Output:
   32 samples. */
/* XXX: optimize by avoiding ring buffer usage */
#if !CONFIG_FLOAT
void ff_mpa_synth_filter(MPA_INT *synth_buf_ptr, int *synth_buf_offset,
                         MPA_INT *window, int *dither_state,
                         OUT_INT *samples, int incr,
                         INTFLOAT sb_samples[SBLIMIT])
{
    register MPA_INT *synth_buf;
    int offset;
#if FRAC_BITS <= 15
    int32_t tmp[32];
    int j;
#endif

    offset = *synth_buf_offset;
    synth_buf = synth_buf_ptr + offset;

#if FRAC_BITS <= 15
    dct32(tmp, sb_samples);
    for(j=0;j<32;j++) {
        /* NOTE: can cause a loss in precision if very high amplitude
           sound */
        synth_buf[j] = av_clip_int16(tmp[j]);
    }
#else
    dct32(synth_buf, sb_samples);
#endif

    apply_window_mp3_c(synth_buf, window, dither_state, samples, incr);

    offset = (offset - 32) & 511;
    *synth_buf_offset = offset;
}
#endif

#define C3 FIXHR(0.86602540378443864676/2)

/* 0.5 / cos(pi*(2*i+1)/36) */
static const INTFLOAT icos36[9] = {
    FIXR(0.50190991877167369479),
    FIXR(0.51763809020504152469), //0
    FIXR(0.55168895948124587824),
    FIXR(0.61038729438072803416),
    FIXR(0.70710678118654752439), //1
    FIXR(0.87172339781054900991),
    FIXR(1.18310079157624925896),
    FIXR(1.93185165257813657349), //2
    FIXR(5.73685662283492756461),
};

/* 0.5 / cos(pi*(2*i+1)/36) */
static const INTFLOAT icos36h[9] = {
    FIXHR(0.50190991877167369479/2),
    FIXHR(0.51763809020504152469/2), //0
    FIXHR(0.55168895948124587824/2),
    FIXHR(0.61038729438072803416/2),
    FIXHR(0.70710678118654752439/2), //1
    FIXHR(0.87172339781054900991/2),
    FIXHR(1.18310079157624925896/4),
    FIXHR(1.93185165257813657349/4), //2
//    FIXHR(5.73685662283492756461),
};

/* 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, icos36h[4], 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, icos36h[1], 1);
    out[ 8]=
    out[ 9]= in4 + in1;
    out[ 2]=
    out[ 3]= in4 - in1;

    in0 -= in2;
    in5 = MULH3(in5 - in3, icos36h[7], 2);
    out[ 0]=
    out[ 5]= in0 - in5;
    out[ 6]=
    out[11]= in0 + in5;
}

/* cos(pi*i/18) */
#define C1 FIXHR(0.98480775301220805936/2)
#define C2 FIXHR(0.93969262078590838405/2)
#define C3 FIXHR(0.86602540378443864676/2)
#define C4 FIXHR(0.76604444311897803520/2)
#define C5 FIXHR(0.64278760968653932632/2)
#define C6 FIXHR(0.5/2)
#define C7 FIXHR(0.34202014332566873304/2)
#define C8 FIXHR(0.17364817766693034885/2)


/* using Lee like decomposition followed by hand coded 9 points DCT */
static void imdct36(INTFLOAT *out, INTFLOAT *buf, INTFLOAT *in, INTFLOAT *win)
{
    int i, j;
    INTFLOAT t0, t1, t2, t3, s0, s1, s2, s3;
    INTFLOAT tmp[18], *tmp1, *in1;

    for(i=17;i>=1;i--)
        in[i] += in[i-1];
    for(i=17;i>=3;i-=2)
        in[i] += in[i-2];

    for(j=0;j<2;j++) {
        tmp1 = tmp + j;
        in1 = in + j;

        t2 = in1[2*4] + in1[2*8] - in1[2*2];

        t3 = in1[2*0] + SHR(in1[2*6],1);
        t1 = in1[2*0] - in1[2*6];
        tmp1[ 6] = t1 - SHR(t2,1);
        tmp1[16] = t1 + t2;

        t0 = MULH3(in1[2*2] + in1[2*4] ,    C2, 2);
        t1 = MULH3(in1[2*4] - in1[2*8] , -2*C8, 1);
        t2 = MULH3(in1[2*2] + in1[2*8] ,   -C4, 2);

        tmp1[10] = t3 - t0 - t2;
        tmp1[ 2] = t3 + t0 + t1;
        tmp1[14] = t3 + t2 - t1;

        tmp1[ 4] = MULH3(in1[2*5] + in1[2*7] - in1[2*1], -C3, 2);
        t2 = MULH3(in1[2*1] + in1[2*5],    C1, 2);
        t3 = MULH3(in1[2*5] - in1[2*7], -2*C7, 1);
        t0 = MULH3(in1[2*3], C3, 2);

        t1 = MULH3(in1[2*1] + in1[2*7],   -C5, 2);

        tmp1[ 0] = t2 + t3 + t0;
        tmp1[12] = t2 + t1 - t0;
        tmp1[ 8] = t3 - t1 - t0;
    }

    i = 0;
    for(j=0;j<4;j++) {
        t0 = tmp[i];
        t1 = tmp[i + 2];
        s0 = t1 + t0;
        s2 = t1 - t0;

        t2 = tmp[i + 1];
        t3 = tmp[i + 3];
        s1 = MULH3(t3 + t2, icos36h[j], 2);
        s3 = MULLx(t3 - t2, icos36[8 - j], FRAC_BITS);

        t0 = s0 + s1;
        t1 = s0 - s1;
        out[(9 + j)*SBLIMIT] =  MULH3(t1, win[9 + j], 1) + buf[9 + j];
        out[(8 - j)*SBLIMIT] =  MULH3(t1, win[8 - j], 1) + buf[8 - j];
        buf[9 + j] = MULH3(t0, win[18 + 9 + j], 1);
        buf[8 - j] = MULH3(t0, win[18 + 8 - j], 1);

        t0 = s2 + s3;
        t1 = s2 - s3;
        out[(9 + 8 - j)*SBLIMIT] =  MULH3(t1, win[9 + 8 - j], 1) + buf[9 + 8 - j];
        out[(        j)*SBLIMIT] =  MULH3(t1, win[        j], 1) + buf[        j];
        buf[9 + 8 - j] = MULH3(t0, win[18 + 9 + 8 - j], 1);
        buf[      + j] = MULH3(t0, win[18         + j], 1);
        i += 4;
    }

    s0 = tmp[16];
    s1 = MULH3(tmp[17], icos36h[4], 2);
    t0 = s0 + s1;
    t1 = s0 - s1;
    out[(9 + 4)*SBLIMIT] =  MULH3(t1, win[9 + 4], 1) + buf[9 + 4];
    out[(8 - 4)*SBLIMIT] =  MULH3(t1, win[8 - 4], 1) + buf[8 - 4];
    buf[9 + 4] = MULH3(t0, win[18 + 9 + 4], 1);
    buf[8 - 4] = MULH3(t0, win[18 + 8 - 4], 1);
}

/* 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;

    dprintf(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) {
                            /* 3 values at the same time */
                            v = get_bits(&s->gb, -bits);
                            steps = ff_mpa_quant_steps[qindex];
                            s->sb_samples[ch][k * 12 + l + 0][i] =
                                l2_unscale_group(steps, v % steps, scale);
                            v = v / steps;
                            s->sb_samples[ch][k * 12 + l + 1][i] =
                                l2_unscale_group(steps, v % steps, scale);
                            v = v / steps;
                            s->sb_samples[ch][k * 12 + l + 2][i] =
                                l2_unscale_group(steps, v, 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)
{
    if (n == 0)
        return 0;
    else
        return get_bits(s, n);
}


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){
//                av_log(NULL, AV_LOG_ERROR, "pos: %d %d %d %d\n", pos, end_pos, end_pos2, s_index);
                switch_buffer(s, &pos, &end_pos, &end_pos2);
//                av_log(NULL, AV_LOG_ERROR, "new pos: %d %d\n", pos, end_pos);
                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];

            dprintf(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->error_recognition >= FF_ER_COMPLIANT)
                    s_index=0;
                break;
            }
//                av_log(NULL, AV_LOG_ERROR, "pos2: %d %d %d %d\n", pos, end_pos, end_pos2, s_index);
            switch_buffer(s, &pos, &end_pos, &end_pos2);
//                av_log(NULL, AV_LOG_ERROR, "new pos2: %d %d %d\n", pos, end_pos, s_index);
            if(pos >= end_pos)
                break;
        }
        last_pos= pos;

        code = get_vlc2(&s->gb, vlc->table, vlc->bits, 1);
        dprintf(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);
//av_log(NULL, AV_LOG_ERROR, "left:%d buf:%p\n", bits_left, s->in_gb.buffer);
    if (bits_left < 0 && s->error_recognition >= FF_ER_COMPLIANT) {
        av_log(s->avctx, AV_LOG_ERROR, "bits_left=%d\n", bits_left);
        s_index=0;
    }else if(bits_left > 0 && s->error_recognition >= FF_ER_AGGRESSIVE){
        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 + 48;
        }
    } else {
        ptr = g->sb_hybrid;
    }

    for(i=g->short_start;i<13;i++) {
        len = band_size_short[s->sample_rate_index][i];
        ptr1 = ptr;
        dst = tmp;
        for(j=len;j>0;j--) {
            *dst++ = ptr[0*len];
            *dst++ = ptr[1*len];
            *dst++ = ptr[2*len];
            ptr++;
        }
        ptr+=2*len;
        memcpy(ptr1, tmp, len * 3 * sizeof(*ptr1));
    }
}

#define ISQRT2 FIXR(0.70710678118654752440)

static void compute_stereo(MPADecodeContext *s,
                           GranuleDef *g0, GranuleDef *g1)
{
    int i, j, k, l;
    int sf_max, sf, len, non_zero_found;
    INTFLOAT (*is_tab)[16], *tab0, *tab1, tmp0, tmp1, v1, v2;
    int non_zero_found_short[3];

    /* intensity stereo */
    if (s->mode_ext & MODE_EXT_I_STEREO) {
        if (!s->lsf) {
            is_tab = is_table;
            sf_max = 7;
        } else {
            is_tab = is_table_lsf[g1->scalefac_compress & 1];
            sf_max = 16;
        }

        tab0 = g0->sb_hybrid + 576;
        tab1 = g1->sb_hybrid + 576;

        non_zero_found_short[0] = 0;
        non_zero_found_short[1] = 0;
        non_zero_found_short[2] = 0;
        k = (13 - g1->short_start) * 3 + g1->long_end - 3;
        for(i = 12;i >= g1->short_start;i--) {
            /* for last band, use previous scale factor */
            if (i != 11)
                k -= 3;
            len = band_size_short[s->sample_rate_index][i];
            for(l=2;l>=0;l--) {
                tab0 -= len;
                tab1 -= len;
                if (!non_zero_found_short[l]) {
                    /* test if non zero band. if so, stop doing i-stereo */
                    for(j=0;j<len;j++) {
                        if (tab1[j] != 0) {
                            non_zero_found_short[l] = 1;
                            goto found1;
                        }
                    }
                    sf = g1->scale_factors[k + l];
                    if (sf >= sf_max)
                        goto found1;

                    v1 = is_tab[0][sf];
                    v2 = is_tab[1][sf];
                    for(j=0;j<len;j++) {
                        tmp0 = tab0[j];
                        tab0[j] = MULLx(tmp0, v1, FRAC_BITS);
                        tab1[j] = MULLx(tmp0, v2, FRAC_BITS);
                    }
                } else {
                found1:
                    if (s->mode_ext & MODE_EXT_MS_STEREO) {
                        /* lower part of the spectrum : do ms stereo
                           if enabled */
                        for(j=0;j<len;j++) {
                            tmp0 = tab0[j];
                            tmp1 = tab1[j];
                            tab0[j] = MULLx(tmp0 + tmp1, ISQRT2, FRAC_BITS);
                            tab1[j] = MULLx(tmp0 - tmp1, ISQRT2, FRAC_BITS);
                        }
                    }
                }
            }
        }

        non_zero_found = non_zero_found_short[0] |
            non_zero_found_short[1] |
            non_zero_found_short[2];

        for(i = g1->long_end - 1;i >= 0;i--) {
            len = band_size_long[s->sample_rate_index][i];
            tab0 -= len;
            tab1 -= len;
            /* test if non zero band. if so, stop doing i-stereo */
            if (!non_zero_found) {
                for(j=0;j<len;j++) {
                    if (tab1[j] != 0) {
                        non_zero_found = 1;
                        goto found2;
                    }
                }
                /* for last band, use previous scale factor */
                k = (i == 21) ? 20 : i;
                sf = g1->scale_factors[k];
                if (sf >= sf_max)
                    goto found2;
                v1 = is_tab[0][sf];
                v2 = is_tab[1][sf];
                for(j=0;j<len;j++) {
                    tmp0 = tab0[j];
                    tab0[j] = MULLx(tmp0, v1, FRAC_BITS);
                    tab1[j] = MULLx(tmp0, v2, FRAC_BITS);
                }
            } else {
            found2:
                if (s->mode_ext & MODE_EXT_MS_STEREO) {
                    /* lower part of the spectrum : do ms stereo
                       if enabled */
                    for(j=0;j<len;j++) {
                        tmp0 = tab0[j];
                        tmp1 = tab1[j];
                        tab0[j] = MULLx(tmp0 + tmp1, ISQRT2, FRAC_BITS);
                        tab1[j] = MULLx(tmp0 - tmp1, ISQRT2, FRAC_BITS);
                    }
                }
            }
        }
    } else if (s->mode_ext & MODE_EXT_MS_STEREO) {
        /* ms stereo ONLY */
        /* NOTE: the 1/sqrt(2) normalization factor is included in the
           global gain */
        tab0 = g0->sb_hybrid;
        tab1 = g1->sb_hybrid;
        for(i=0;i<576;i++) {
            tmp0 = tab0[i];
            tmp1 = tab1[i];
            tab0[i] = tmp0 + tmp1;
            tab1[i] = tmp0 - tmp1;
        }
    }
}

static void compute_antialias_integer(MPADecodeContext *s,
                              GranuleDef *g)
{
    int32_t *ptr, *csa;
    int n, i;

    /* we antialias only "long" bands */
    if (g->block_type == 2) {
        if (!g->switch_point)
            return;
        /* XXX: check this for 8000Hz case */
        n = 1;
    } else {
        n = SBLIMIT - 1;
    }

    ptr = g->sb_hybrid + 18;
    for(i = n;i > 0;i--) {
        int tmp0, tmp1, tmp2;
        csa = &csa_table[0][0];
#define INT_AA(j) \
            tmp0 = ptr[-1-j];\
            tmp1 = ptr[   j];\
            tmp2= MULH(tmp0 + tmp1, csa[0+4*j]);\
            ptr[-1-j] = 4*(tmp2 - MULH(tmp1, csa[2+4*j]));\
            ptr[   j] = 4*(tmp2 + MULH(tmp0, csa[3+4*j]));

        INT_AA(0)
        INT_AA(1)
        INT_AA(2)
        INT_AA(3)
        INT_AA(4)
        INT_AA(5)
        INT_AA(6)
        INT_AA(7)

        ptr += 18;
    }
}

static void compute_antialias_float(MPADecodeContext *s,
                              GranuleDef *g)
{
    float *ptr;
    int n, i;

    /* we antialias only "long" bands */
    if (g->block_type == 2) {
        if (!g->switch_point)
            return;
        /* XXX: check this for 8000Hz case */
        n = 1;
    } else {
        n = SBLIMIT - 1;
    }

    ptr = g->sb_hybrid + 18;
    for(i = n;i > 0;i--) {
        float tmp0, tmp1;
        float *csa = &csa_table_float[0][0];
#define FLOAT_AA(j)\
        tmp0= ptr[-1-j];\
        tmp1= ptr[   j];\
        ptr[-1-j] = tmp0 * csa[0+4*j] - tmp1 * csa[1+4*j];\
        ptr[   j] = tmp0 * csa[1+4*j] + tmp1 * csa[0+4*j];

        FLOAT_AA(0)
        FLOAT_AA(1)
        FLOAT_AA(2)
        FLOAT_AA(3)
        FLOAT_AA(4)
        FLOAT_AA(5)
        FLOAT_AA(6)
        FLOAT_AA(7)

        ptr += 18;
    }
}

static void compute_imdct(MPADecodeContext *s,
                          GranuleDef *g,
                          INTFLOAT *sb_samples,
                          INTFLOAT *mdct_buf)
{
    INTFLOAT *win, *win1, *out_ptr, *ptr, *buf, *ptr1;
    INTFLOAT out2[12];
    int i, j, mdct_long_end, sblimit;

    /* find last non zero block */
    ptr = g->sb_hybrid + 576;
    ptr1 = g->sb_hybrid + 2 * 18;
    while (ptr >= ptr1) {
        int32_t *p;
        ptr -= 6;
        p= (int32_t*)ptr;
        if(p[0] | p[1] | p[2] | p[3] | p[4] | p[5])
            break;
    }
    sblimit = ((ptr - g->sb_hybrid) / 18) + 1;

    if (g->block_type == 2) {
        /* XXX: check for 8000 Hz */
        if (g->switch_point)
            mdct_long_end = 2;
        else
            mdct_long_end = 0;
    } else {
        mdct_long_end = sblimit;
    }

    buf = mdct_buf;
    ptr = g->sb_hybrid;
    for(j=0;j<mdct_long_end;j++) {
        /* apply window & overlap with previous buffer */
        out_ptr = sb_samples + j;
        /* select window */
        if (g->switch_point && j < 2)
            win1 = mdct_win[0];
        else
            win1 = mdct_win[g->block_type];
        /* select frequency inversion */
        win = win1 + ((4 * 36) & -(j & 1));
        imdct36(out_ptr, buf, ptr, win);
        out_ptr += 18*SBLIMIT;
        ptr += 18;
        buf += 18;
    }
    for(j=mdct_long_end;j<sblimit;j++) {
        /* select frequency inversion */
        win = mdct_win[2] + ((4 * 36) & -(j & 1));
        out_ptr = sb_samples + j;

        for(i=0; i<6; i++){
            *out_ptr = buf[i];
            out_ptr += SBLIMIT;
        }
        imdct12(out2, ptr + 0);
        for(i=0;i<6;i++) {
            *out_ptr     = MULH3(out2[i    ], win[i    ], 1) + buf[i + 6*1];
            buf[i + 6*2] = MULH3(out2[i + 6], win[i + 6], 1);
            out_ptr += SBLIMIT;
        }
        imdct12(out2, ptr + 1);
        for(i=0;i<6;i++) {
            *out_ptr     = MULH3(out2[i    ], win[i    ], 1) + buf[i + 6*2];
            buf[i + 6*0] = MULH3(out2[i + 6], win[i + 6], 1);
            out_ptr += SBLIMIT;
        }
        imdct12(out2, ptr + 2);
        for(i=0;i<6;i++) {
            buf[i + 6*0] = MULH3(out2[i    ], win[i    ], 1) + buf[i + 6*0];
            buf[i + 6*1] = MULH3(out2[i + 6], win[i + 6], 1);
            buf[i + 6*2] = 0;
        }
        ptr += 18;
        buf += 18;
    }
    /* zero bands */
    for(j=sblimit;j<SBLIMIT;j++) {
        /* overlap */
        out_ptr = sb_samples + j;
        for(i=0;i<18;i++) {
            *out_ptr = buf[i];
            buf[i] = 0;
            out_ptr += SBLIMIT;
        }
        buf += 18;
    }
}

/* main layer3 decoding function */
static int mp_decode_layer3(MPADecodeContext *s)
{
    int nb_granules, main_data_begin, private_bits;
    int gr, ch, blocksplit_flag, i, j, k, n, bits_pos;
    GranuleDef *g;
    int16_t exponents[576]; //FIXME try INTFLOAT

    /* read side info */
    if (s->lsf) {
        main_data_begin = get_bits(&s->gb, 8);
        private_bits = get_bits(&s->gb, s->nb_channels);
        nb_granules = 1;
    } else {
        main_data_begin = get_bits(&s->gb, 9);
        if (s->nb_channels == 2)
            private_bits = get_bits(&s->gb, 3);
        else
            private_bits = get_bits(&s->gb, 5);
        nb_granules = 2;
        for(ch=0;ch<s->nb_channels;ch++) {
            s->granules[ch][0].scfsi = 0;/* all scale factors are transmitted */
            s->granules[ch][1].scfsi = get_bits(&s->gb, 4);
        }
    }

    for(gr=0;gr<nb_granules;gr++) {
        for(ch=0;ch<s->nb_channels;ch++) {
            dprintf(s->avctx, "gr=%d ch=%d: side_info\n", gr, ch);
            g = &s->granules[ch][gr];
            g->part2_3_length = get_bits(&s->gb, 12);
            g->big_values = get_bits(&s->gb, 9);
            if(g->big_values > 288){
                av_log(s->avctx, AV_LOG_ERROR, "big_values too big\n");
                return -1;
            }

            g->global_gain = get_bits(&s->gb, 8);
            /* if MS stereo only is selected, we precompute the
               1/sqrt(2) renormalization factor */
            if ((s->mode_ext & (MODE_EXT_MS_STEREO | MODE_EXT_I_STEREO)) ==
                MODE_EXT_MS_STEREO)
                g->global_gain -= 2;
            if (s->lsf)
                g->scalefac_compress = get_bits(&s->gb, 9);
            else
                g->scalefac_compress = get_bits(&s->gb, 4);
            blocksplit_flag = get_bits1(&s->gb);
            if (blocksplit_flag) {
                g->block_type = get_bits(&s->gb, 2);
                if (g->block_type == 0){
                    av_log(s->avctx, AV_LOG_ERROR, "invalid block type\n");
                    return -1;
                }
                g->switch_point = get_bits1(&s->gb);
                for(i=0;i<2;i++)
                    g->table_select[i] = get_bits(&s->gb, 5);
                for(i=0;i<3;i++)
                    g->subblock_gain[i] = get_bits(&s->gb, 3);
                ff_init_short_region(s, g);
            } else {
                int region_address1, region_address2;
                g->block_type = 0;
                g->switch_point = 0;
                for(i=0;i<3;i++)
                    g->table_select[i] = get_bits(&s->gb, 5);
                /* compute huffman coded region sizes */
                region_address1 = get_bits(&s->gb, 4);
                region_address2 = get_bits(&s->gb, 3);
                dprintf(s->avctx, "region1=%d region2=%d\n",
                        region_address1, region_address2);
                ff_init_long_region(s, g, region_address1, region_address2);
            }
            ff_region_offset2size(g);
            ff_compute_band_indexes(s, g);

            g->preflag = 0;
            if (!s->lsf)
                g->preflag = get_bits1(&s->gb);
            g->scalefac_scale = get_bits1(&s->gb);
            g->count1table_select = get_bits1(&s->gb);
            dprintf(s->avctx, "block_type=%d switch_point=%d\n",
                    g->block_type, g->switch_point);
        }
    }

  if (!s->adu_mode) {
    const uint8_t *ptr = s->gb.buffer + (get_bits_count(&s->gb)>>3);
    assert((get_bits_count(&s->gb) & 7) == 0);
    /* now we get bits from the main_data_begin offset */
    dprintf(s->avctx, "seekback: %d\n", main_data_begin);
//av_log(NULL, AV_LOG_ERROR, "backstep:%d, lastbuf:%d\n", main_data_begin, s->last_buf_size);

    memcpy(s->last_buf + s->last_buf_size, ptr, EXTRABYTES);
    s->in_gb= s->gb;
        init_get_bits(&s->gb, s->last_buf, s->last_buf_size*8);
        skip_bits_long(&s->gb, 8*(s->last_buf_size - main_data_begin));
  }

    for(gr=0;gr<nb_granules;gr++) {
        for(ch=0;ch<s->nb_channels;ch++) {
            g = &s->granules[ch][gr];
            if(get_bits_count(&s->gb)<0){
                av_log(s->avctx, AV_LOG_DEBUG, "mdb:%d, lastbuf:%d skipping granule %d\n",
                                            main_data_begin, s->last_buf_size, gr);
                skip_bits_long(&s->gb, g->part2_3_length);
                memset(g->sb_hybrid, 0, sizeof(g->sb_hybrid));
                if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->in_gb.buffer){
                    skip_bits_long(&s->in_gb, get_bits_count(&s->gb) - s->gb.size_in_bits);
                    s->gb= s->in_gb;
                    s->in_gb.buffer=NULL;
                }
                continue;
            }

            bits_pos = get_bits_count(&s->gb);

            if (!s->lsf) {
                uint8_t *sc;
                int slen, slen1, slen2;

                /* MPEG1 scale factors */
                slen1 = slen_table[0][g->scalefac_compress];
                slen2 = slen_table[1][g->scalefac_compress];
                dprintf(s->avctx, "slen1=%d slen2=%d\n", slen1, slen2);
                if (g->block_type == 2) {
                    n = g->switch_point ? 17 : 18;
                    j = 0;
                    if(slen1){
                        for(i=0;i<n;i++)
                            g->scale_factors[j++] = get_bits(&s->gb, slen1);
                    }else{
                        for(i=0;i<n;i++)
                            g->scale_factors[j++] = 0;
                    }
                    if(slen2){
                        for(i=0;i<18;i++)
                            g->scale_factors[j++] = get_bits(&s->gb, slen2);
                        for(i=0;i<3;i++)
                            g->scale_factors[j++] = 0;
                    }else{
                        for(i=0;i<21;i++)
                            g->scale_factors[j++] = 0;
                    }
                } else {
                    sc = s->granules[ch][0].scale_factors;
                    j = 0;
                    for(k=0;k<4;k++) {
                        n = (k == 0 ? 6 : 5);
                        if ((g->scfsi & (0x8 >> k)) == 0) {
                            slen = (k < 2) ? slen1 : slen2;
                            if(slen){
                                for(i=0;i<n;i++)
                                    g->scale_factors[j++] = get_bits(&s->gb, slen);
                            }else{
                                for(i=0;i<n;i++)
                                    g->scale_factors[j++] = 0;
                            }
                        } else {
                            /* simply copy from last granule */
                            for(i=0;i<n;i++) {
                                g->scale_factors[j] = sc[j];
                                j++;
                            }
                        }
                    }
                    g->scale_factors[j++] = 0;
                }
            } else {
                int tindex, tindex2, slen[4], sl, sf;

                /* LSF scale factors */
                if (g->block_type == 2) {
                    tindex = g->switch_point ? 2 : 1;
                } else {
                    tindex = 0;
                }
                sf = g->scalefac_compress;
                if ((s->mode_ext & MODE_EXT_I_STEREO) && ch == 1) {
                    /* intensity stereo case */
                    sf >>= 1;
                    if (sf < 180) {
                        lsf_sf_expand(slen, sf, 6, 6, 0);
                        tindex2 = 3;
                    } else if (sf < 244) {
                        lsf_sf_expand(slen, sf - 180, 4, 4, 0);
                        tindex2 = 4;
                    } else {
                        lsf_sf_expand(slen, sf - 244, 3, 0, 0);
                        tindex2 = 5;
                    }
                } else {
                    /* normal case */
                    if (sf < 400) {
                        lsf_sf_expand(slen, sf, 5, 4, 4);
                        tindex2 = 0;
                    } else if (sf < 500) {
                        lsf_sf_expand(slen, sf - 400, 5, 4, 0);
                        tindex2 = 1;
                    } else {
                        lsf_sf_expand(slen, sf - 500, 3, 0, 0);
                        tindex2 = 2;
                        g->preflag = 1;
                    }
                }

                j = 0;
                for(k=0;k<4;k++) {
                    n = lsf_nsf_table[tindex2][tindex][k];
                    sl = slen[k];
                    if(sl){
                        for(i=0;i<n;i++)
                            g->scale_factors[j++] = get_bits(&s->gb, sl);
                    }else{
                        for(i=0;i<n;i++)
                            g->scale_factors[j++] = 0;
                    }
                }
                /* XXX: should compute exact size */
                for(;j<40;j++)
                    g->scale_factors[j] = 0;
            }

            exponents_from_scale_factors(s, g, exponents);

            /* read Huffman coded residue */
            huffman_decode(s, g, exponents, bits_pos + g->part2_3_length);
        } /* ch */

        if (s->nb_channels == 2)
            compute_stereo(s, &s->granules[0][gr], &s->granules[1][gr]);

        for(ch=0;ch<s->nb_channels;ch++) {
            g = &s->granules[ch][gr];

            reorder_block(s, g);
            compute_antialias(s, g);
            compute_imdct(s, g, &s->sb_samples[ch][18 * gr][0], s->mdct_buf[ch]);
        }
    } /* gr */
    if(get_bits_count(&s->gb)<0)
        skip_bits_long(&s->gb, -get_bits_count(&s->gb));
    return nb_granules * 18;
}

static int mp_decode_frame(MPADecodeContext *s,
                           OUT_INT *samples, const uint8_t *buf, int buf_size)
{
    int i, nb_frames, ch;
    OUT_INT *samples_ptr;

    init_get_bits(&s->gb, buf + HEADER_SIZE, (buf_size - HEADER_SIZE)*8);

    /* skip error protection field */
    if (s->error_protection)
        skip_bits(&s->gb, 16);

    dprintf(s->avctx, "frame %d:\n", s->frame_count);
    switch(s->layer) {
    case 1:
        s->avctx->frame_size = 384;
        nb_frames = mp_decode_layer1(s);
        break;
    case 2:
        s->avctx->frame_size = 1152;
        nb_frames = mp_decode_layer2(s);
        break;
    case 3:
        s->avctx->frame_size = s->lsf ? 576 : 1152;
    default:
        nb_frames = mp_decode_layer3(s);

        s->last_buf_size=0;
        if(s->in_gb.buffer){
            align_get_bits(&s->gb);
            i= get_bits_left(&s->gb)>>3;
            if(i >= 0 && i <= BACKSTEP_SIZE){
                memmove(s->last_buf, s->gb.buffer + (get_bits_count(&s->gb)>>3), i);
                s->last_buf_size=i;
            }else
                av_log(s->avctx, AV_LOG_ERROR, "invalid old backstep %d\n", i);
            s->gb= s->in_gb;
            s->in_gb.buffer= NULL;
        }

        align_get_bits(&s->gb);
        assert((get_bits_count(&s->gb) & 7) == 0);
        i= get_bits_left(&s->gb)>>3;

        if(i<0 || i > BACKSTEP_SIZE || nb_frames<0){
            if(i<0)
                av_log(s->avctx, AV_LOG_ERROR, "invalid new backstep %d\n", i);
            i= FFMIN(BACKSTEP_SIZE, buf_size - HEADER_SIZE);
        }
        assert(i <= buf_size - HEADER_SIZE && i>= 0);
        memcpy(s->last_buf + s->last_buf_size, s->gb.buffer + buf_size - HEADER_SIZE - i, i);
        s->last_buf_size += i;

        break;
    }

    /* apply the synthesis filter */
    for(ch=0;ch<s->nb_channels;ch++) {
        samples_ptr = samples + ch;
        for(i=0;i<nb_frames;i++) {
            RENAME(ff_mpa_synth_filter)(
#if CONFIG_FLOAT
                         s,
#endif
                         s->synth_buf[ch], &(s->synth_buf_offset[ch]),
                         RENAME(ff_mpa_synth_window), &s->dither_state,
                         samples_ptr, s->nb_channels,
                         s->sb_samples[ch][i]);
            samples_ptr += 32 * s->nb_channels;
        }
    }

    return nb_frames * 32 * sizeof(OUT_INT) * s->nb_channels;
}

static int decode_frame(AVCodecContext * avctx,
                        void *data, int *data_size,
                        AVPacket *avpkt)
{
    const uint8_t *buf = avpkt->data;
    int buf_size = avpkt->size;
    MPADecodeContext *s = avctx->priv_data;
    uint32_t header;
    int out_size;
    OUT_INT *out_samples = data;

    if(buf_size < HEADER_SIZE)
        return -1;

    header = AV_RB32(buf);
    if(ff_mpa_check_header(header) < 0){
        av_log(avctx, AV_LOG_ERROR, "Header missing\n");
        return -1;
    }

    if (ff_mpegaudio_decode_header((MPADecodeHeader *)s, header) == 1) {
        /* free format: prepare to compute frame size */
        s->frame_size = -1;
        return -1;
    }
    /* update codec info */
    avctx->channels = s->nb_channels;
    avctx->bit_rate = s->bit_rate;
    avctx->sub_id = s->layer;

    if(*data_size < 1152*avctx->channels*sizeof(OUT_INT))
        return -1;
    *data_size = 0;

    if(s->frame_size<=0 || s->frame_size > buf_size){
        av_log(avctx, AV_LOG_ERROR, "incomplete frame\n");
        return -1;
    }else if(s->frame_size < buf_size){
        av_log(avctx, AV_LOG_ERROR, "incorrect frame size\n");
        buf_size= s->frame_size;
    }

    out_size = mp_decode_frame(s, out_samples, buf, buf_size);
    if(out_size>=0){
        *data_size = out_size;
        avctx->sample_rate = s->sample_rate;
        //FIXME maybe move the other codec info stuff from above here too
    }else
        av_log(avctx, AV_LOG_DEBUG, "Error while decoding MPEG audio frame.\n"); //FIXME return -1 / but also return the number of bytes consumed
    s->frame_size = 0;
    return buf_size;
}

static void flush(AVCodecContext *avctx){
    MPADecodeContext *s = avctx->priv_data;
    memset(s->synth_buf, 0, sizeof(s->synth_buf));
    s->last_buf_size= 0;
}

#if CONFIG_MP3ADU_DECODER
static int decode_frame_adu(AVCodecContext * avctx,
                        void *data, int *data_size,
                        AVPacket *avpkt)
{
    const uint8_t *buf = avpkt->data;
    int buf_size = avpkt->size;
    MPADecodeContext *s = avctx->priv_data;
    uint32_t header;
    int len, out_size;
    OUT_INT *out_samples = data;

    len = buf_size;

    // Discard too short frames
    if (buf_size < HEADER_SIZE) {
        *data_size = 0;
        return buf_size;
    }


    if (len > MPA_MAX_CODED_FRAME_SIZE)
        len = MPA_MAX_CODED_FRAME_SIZE;

    // Get header and restore sync word
    header = AV_RB32(buf) | 0xffe00000;

    if (ff_mpa_check_header(header) < 0) { // Bad header, discard frame
        *data_size = 0;
        return buf_size;
    }

    ff_mpegaudio_decode_header((MPADecodeHeader *)s, header);
    /* update codec info */
    avctx->sample_rate = s->sample_rate;
    avctx->channels = s->nb_channels;
    avctx->bit_rate = s->bit_rate;
    avctx->sub_id = s->layer;

    s->frame_size = len;

    if (avctx->parse_only) {
        out_size = buf_size;
    } else {
        out_size = mp_decode_frame(s, out_samples, buf, buf_size);
    }

    *data_size = out_size;
    return buf_size;
}
#endif /* CONFIG_MP3ADU_DECODER */

#if CONFIG_MP3ON4_DECODER

/**
 * Context for MP3On4 decoder
 */
typedef struct MP3On4DecodeContext {
    int frames;   ///< number of mp3 frames per block (number of mp3 decoder instances)
    int syncword; ///< syncword patch
    const uint8_t *coff; ///< channels offsets in output buffer
    MPADecodeContext *mp3decctx[5]; ///< MPADecodeContext for every decoder instance
} MP3On4DecodeContext;

#include "mpeg4audio.h"

/* Next 3 arrays are indexed by channel config number (passed via codecdata) */
static const uint8_t mp3Frames[8] = {0,1,1,2,3,3,4,5};   /* number of mp3 decoder instances */
/* offsets into output buffer, assume output order is FL FR BL BR C LFE */
static const uint8_t chan_offset[8][5] = {
    {0},
    {0},            // C
    {0},            // FLR
    {2,0},          // C FLR
    {2,0,3},        // C FLR BS
    {4,0,2},        // C FLR BLRS
    {4,0,2,5},      // C FLR BLRS LFE
    {4,0,2,6,5},    // C FLR BLRS BLR LFE
};


static int decode_init_mp3on4(AVCodecContext * avctx)
{
    MP3On4DecodeContext *s = avctx->priv_data;
    MPEG4AudioConfig cfg;
    int i;

    if ((avctx->extradata_size < 2) || (avctx->extradata == NULL)) {
        av_log(avctx, AV_LOG_ERROR, "Codec extradata missing or too short.\n");
        return -1;
    }

    ff_mpeg4audio_get_config(&cfg, avctx->extradata, avctx->extradata_size);
    if (!cfg.chan_config || cfg.chan_config > 7) {
        av_log(avctx, AV_LOG_ERROR, "Invalid channel config number.\n");
        return -1;
    }
    s->frames = mp3Frames[cfg.chan_config];
    s->coff = chan_offset[cfg.chan_config];
    avctx->channels = ff_mpeg4audio_channels[cfg.chan_config];

    if (cfg.sample_rate < 16000)
        s->syncword = 0xffe00000;
    else
        s->syncword = 0xfff00000;

    /* Init the first mp3 decoder in standard way, so that all tables get builded
     * We replace avctx->priv_data with the context of the first decoder so that
     * decode_init() does not have to be changed.
     * Other decoders will be initialized here copying data from the first context
     */
    // Allocate zeroed memory for the first decoder context
    s->mp3decctx[0] = av_mallocz(sizeof(MPADecodeContext));
    // Put decoder context in place to make init_decode() happy
    avctx->priv_data = s->mp3decctx[0];
    decode_init(avctx);
    // Restore mp3on4 context pointer
    avctx->priv_data = s;
    s->mp3decctx[0]->adu_mode = 1; // Set adu mode

    /* Create a separate codec/context for each frame (first is already ok).
     * Each frame is 1 or 2 channels - up to 5 frames allowed
     */
    for (i = 1; i < s->frames; i++) {
        s->mp3decctx[i] = av_mallocz(sizeof(MPADecodeContext));
        s->mp3decctx[i]->adu_mode = 1;
        s->mp3decctx[i]->avctx = avctx;
    }

    return 0;
}


static av_cold int decode_close_mp3on4(AVCodecContext * avctx)
{
    MP3On4DecodeContext *s = avctx->priv_data;
    int i;

    for (i = 0; i < s->frames; i++)
        if (s->mp3decctx[i])
            av_free(s->mp3decctx[i]);

    return 0;
}


static int decode_frame_mp3on4(AVCodecContext * avctx,
                        void *data, int *data_size,
                        AVPacket *avpkt)
{
    const uint8_t *buf = avpkt->data;
    int buf_size = avpkt->size;
    MP3On4DecodeContext *s = avctx->priv_data;
    MPADecodeContext *m;
    int fsize, len = buf_size, out_size = 0;
    uint32_t header;
    OUT_INT *out_samples = data;
    OUT_INT decoded_buf[MPA_FRAME_SIZE * MPA_MAX_CHANNELS];
    OUT_INT *outptr, *bp;
    int fr, j, n;

    if(*data_size < MPA_FRAME_SIZE * MPA_MAX_CHANNELS * s->frames * sizeof(OUT_INT))
        return -1;

    *data_size = 0;
    // Discard too short frames
    if (buf_size < HEADER_SIZE)
        return -1;

    // If only one decoder interleave is not needed
    outptr = s->frames == 1 ? out_samples : decoded_buf;

    avctx->bit_rate = 0;

    for (fr = 0; fr < s->frames; fr++) {
        fsize = AV_RB16(buf) >> 4;
        fsize = FFMIN3(fsize, len, MPA_MAX_CODED_FRAME_SIZE);
        m = s->mp3decctx[fr];
        assert (m != NULL);

        header = (AV_RB32(buf) & 0x000fffff) | s->syncword; // patch header

        if (ff_mpa_check_header(header) < 0) // Bad header, discard block
            break;

        ff_mpegaudio_decode_header((MPADecodeHeader *)m, header);
        out_size += mp_decode_frame(m, outptr, buf, fsize);
        buf += fsize;
        len -= fsize;

        if(s->frames > 1) {
            n = m->avctx->frame_size*m->nb_channels;
            /* interleave output data */
            bp = out_samples + s->coff[fr];
            if(m->nb_channels == 1) {
                for(j = 0; j < n; j++) {
                    *bp = decoded_buf[j];
                    bp += avctx->channels;
                }
            } else {
                for(j = 0; j < n; j++) {
                    bp[0] = decoded_buf[j++];
                    bp[1] = decoded_buf[j];
                    bp += avctx->channels;
                }
            }
        }
        avctx->bit_rate += m->bit_rate;
    }

    /* update codec info */
    avctx->sample_rate = s->mp3decctx[0]->sample_rate;

    *data_size = out_size;
    return buf_size;
}
#endif /* CONFIG_MP3ON4_DECODER */

#if !CONFIG_FLOAT
#if CONFIG_MP1_DECODER
AVCodec mp1_decoder =
{
    "mp1",
    AVMEDIA_TYPE_AUDIO,
    CODEC_ID_MP1,
    sizeof(MPADecodeContext),
    decode_init,
    NULL,
    NULL,
    decode_frame,
    CODEC_CAP_PARSE_ONLY,
    .flush= flush,
    .long_name= NULL_IF_CONFIG_SMALL("MP1 (MPEG audio layer 1)"),
};
#endif
#if CONFIG_MP2_DECODER
AVCodec mp2_decoder =
{
    "mp2",
    AVMEDIA_TYPE_AUDIO,
    CODEC_ID_MP2,
    sizeof(MPADecodeContext),
    decode_init,
    NULL,
    NULL,
    decode_frame,
    CODEC_CAP_PARSE_ONLY,
    .flush= flush,
    .long_name= NULL_IF_CONFIG_SMALL("MP2 (MPEG audio layer 2)"),
};
#endif
#if CONFIG_MP3_DECODER
AVCodec mp3_decoder =
{
    "mp3",
    AVMEDIA_TYPE_AUDIO,
    CODEC_ID_MP3,
    sizeof(MPADecodeContext),
    decode_init,
    NULL,
    NULL,
    decode_frame,
    CODEC_CAP_PARSE_ONLY,
    .flush= flush,
    .long_name= NULL_IF_CONFIG_SMALL("MP3 (MPEG audio layer 3)"),
};
#endif
#if CONFIG_MP3ADU_DECODER
AVCodec mp3adu_decoder =
{
    "mp3adu",
    AVMEDIA_TYPE_AUDIO,
    CODEC_ID_MP3ADU,
    sizeof(MPADecodeContext),
    decode_init,
    NULL,
    NULL,
    decode_frame_adu,
    CODEC_CAP_PARSE_ONLY,
    .flush= flush,
    .long_name= NULL_IF_CONFIG_SMALL("ADU (Application Data Unit) MP3 (MPEG audio layer 3)"),
};
#endif
#if CONFIG_MP3ON4_DECODER
AVCodec mp3on4_decoder =
{
    "mp3on4",
    AVMEDIA_TYPE_AUDIO,
    CODEC_ID_MP3ON4,
    sizeof(MP3On4DecodeContext),
    decode_init_mp3on4,
    NULL,
    decode_close_mp3on4,
    decode_frame_mp3on4,
    .flush= flush,
    .long_name= NULL_IF_CONFIG_SMALL("MP3onMP4"),
};
#endif
#endif