mpegaudiodec.c

/*
 * 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)*(1.0/(1<<(b))))
#   define compute_antialias compute_antialias_float
#   define FIXR_OLD(a)    ((int)((a) * FRAC_ONE + 0.5))
#   define FIXR(x)        (x)
#   define FIXHR(x)       (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);

/* 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];

/**
 * 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;

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

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

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

#define ADD(a, b) tab[a] += tab[b]

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

    /* pass 1 */
    BF( 0, 31, COS0_0 , 1);
    BF(15, 16, COS0_15, 5);
    /* pass 2 */
    BF( 0, 15, COS1_0 , 1);
    BF(16, 31,-COS1_0 , 1);
    /* pass 1 */
    BF( 7, 24, COS0_7 , 1);
    BF( 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 */
    BF( 3, 28, COS0_3 , 1);
    BF(12, 19, COS0_12, 2);
    /* pass 2 */
    BF( 3, 12, COS1_3 , 1);
    BF(19, 28,-COS1_3 , 1);
    /* pass 1 */
    BF( 4, 27, COS0_4 , 1);
    BF(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 */
    BF( 1, 30, COS0_1 , 1);
    BF(14, 17, COS0_14, 3);
    /* pass 2 */
    BF( 1, 14, COS1_1 , 1);
    BF(17, 30,-COS1_1 , 1);
    /* pass 1 */
    BF( 6, 25, COS0_6 , 1);
    BF( 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 */
    BF( 2, 29, COS0_2 , 1);
    BF(13, 18, COS0_13, 3);
    /* pass 2 */
    BF( 2, 13, COS1_2 , 1);
    BF(18, 29,-COS1_2 , 1);
    /* pass 1 */
    BF( 5, 26, COS0_5 , 1);
    BF(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] = tab[0];
    out[16] = tab[1];
    out[ 8] = tab[2];
    out[24] = tab[3];
    out[ 4] = tab[4];
    out[20] = tab[5];
    out[12] = tab[6];
    out[28] = tab[7];
    out[ 2] = tab[8];
    out[18] = tab[9];
    out[10] = tab[10];
    out[26] = tab[11];
    out[ 6] = tab[12];
    out[22] = tab[13];
    out[14] = tab[14];
    out[30] = tab[15];

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

    out[ 1] = tab[16] + tab[24];
    out[17] = tab[17] + tab[25];
    out[ 9] = tab[18] + tab[26];
    out[25] = tab[19] + tab[27];
    out[ 5] = tab[20] + tab[28];
    out[21] = tab[21] + tab[29];
    out[13] = tab[22] + tab[30];
    out[29] = tab[23] + tab[31];
    out[ 3] = tab[24] + tab[20];
    out[19] = tab[25] + tab[21];
    out[11] = tab[26] + tab[22];
    out[27] = tab[27] + tab[23];
    out[ 7] = tab[28] + tab[18];
    out[23] = tab[29] + tab[19];
    out[15] = tab[30] + tab[17];
    out[31] = tab[31];
}

#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;

    /* 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;
    }
}

/* 32 sub band synthesis filter. Input: 32 sub band samples, Output:
   32 samples. */
/* XXX: optimize by avoiding ring buffer usage */
void RENAME(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;
    register const MPA_INT *w, *w2, *p;
    int j, offset;
    OUT_INT *samples2;
#if CONFIG_FLOAT
    float sum, sum2;
#elif FRAC_BITS <= 15
    int32_t tmp[32];
    int sum, sum2;
#else
    int64_t sum, sum2;
#endif

    offset = *synth_buf_offset;
    synth_buf = synth_buf_ptr + offset;

#if FRAC_BITS <= 15
    assert(!CONFIG_FLOAT);
    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

    /* 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;

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

#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;