mpegaudiodec.c

/*
 * MPEG Audio decoder
 * Copyright (c) 2001, 2002 Fabrice Bellard.
 *
 * This library 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 of the License, or (at your option) any later version.
 *
 * This library 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 this library; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @file mpegaudiodec.c
 * MPEG Audio decoder.
 */

//#define DEBUG
#include "avcodec.h"
#include "bitstream.h"
#include "dsputil.h"

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

/* define USE_HIGHPRECISION to have a bit exact (but slower) mpeg
   audio decoder */
#ifdef CONFIG_MPEGAUDIO_HP
#   define USE_HIGHPRECISION
#endif

#include "mpegaudio.h"

#define FRAC_ONE    (1 << FRAC_BITS)

#ifdef ARCH_X86
#   define MULL(ra, rb) \
        ({ int rt, dummy; asm (\
            "imull %3               \n\t"\
            "shrdl %4, %%edx, %%eax \n\t"\
            : "=a"(rt), "=d"(dummy)\
            : "a" (ra), "rm" (rb), "i"(FRAC_BITS));\
         rt; })
#   define MUL64(ra, rb) \
        ({ int64_t rt; asm ("imull %2\n\t" : "=A"(rt) : "a" (ra), "g" (rb)); rt; })
#   define MULH(ra, rb) \
        ({ int rt, dummy; asm ("imull %3\n\t" : "=d"(rt), "=a"(dummy): "a" (ra), "rm" (rb)); rt; })
#elif defined(ARCH_ARMV4L)
#   define MULL(a, b) \
        ({  int lo, hi;\
            asm("smull %0, %1, %2, %3     \n\t"\
                "mov   %0, %0,     lsr %4\n\t"\
                "add   %1, %0, %1, lsl %5\n\t"\
            : "=&r"(lo), "=&r"(hi)\
            : "r"(b), "r"(a), "i"(FRAC_BITS), "i"(32-FRAC_BITS));\
         hi; })
#   define MUL64(a,b) ((int64_t)(a) * (int64_t)(b))
#   define MULH(a, b) ({ int lo, hi; asm ("smull %0, %1, %2, %3" : "=&r"(lo), "=&r"(hi) : "r"(b), "r"(a)); hi; })
#else
#   define MULL(a,b) (((int64_t)(a) * (int64_t)(b)) >> FRAC_BITS)
#   define MUL64(a,b) ((int64_t)(a) * (int64_t)(b))
//#define MULH(a,b) (((int64_t)(a) * (int64_t)(b))>>32) //gcc 3.4 creates an incredibly bloated mess out of this
static always_inline int MULH(int a, int b){
    return ((int64_t)(a) * (int64_t)(b))>>32;
}
#endif
#define FIX(a)   ((int)((a) * FRAC_ONE))
/* WARNING: only correct for posititive numbers */
#define FIXR(a)   ((int)((a) * FRAC_ONE + 0.5))
#define FRAC_RND(a) (((a) + (FRAC_ONE/2)) >> FRAC_BITS)

#define FIXHR(a) ((int)((a) * (1LL<<32) + 0.5))

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

#define HEADER_SIZE 4
#define BACKSTEP_SIZE 512

struct GranuleDef;

typedef struct MPADecodeContext {
    uint8_t inbuf1[2][MPA_MAX_CODED_FRAME_SIZE + BACKSTEP_SIZE];        /* input buffer */
    int inbuf_index;
    uint8_t *inbuf_ptr, *inbuf;
    int frame_size;
    int free_format_frame_size; /* frame size in case of free format
                                   (zero if currently unknown) */
    /* next header (used in free format parsing) */
    uint32_t free_format_next_header;
    int error_protection;
    int layer;
    int sample_rate;
    int sample_rate_index; /* between 0 and 8 */
    int bit_rate;
    int old_frame_size;
    GetBitContext gb;
    int nb_channels;
    int mode;
    int mode_ext;
    int lsf;
    MPA_INT synth_buf[MPA_MAX_CHANNELS][512 * 2] __attribute__((aligned(16)));
    int synth_buf_offset[MPA_MAX_CHANNELS];
    int32_t sb_samples[MPA_MAX_CHANNELS][36][SBLIMIT] __attribute__((aligned(16)));
    int32_t mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
#ifdef DEBUG
    int frame_count;
#endif
    void (*compute_antialias)(struct MPADecodeContext *s, struct GranuleDef *g);
    int adu_mode; ///< 0 for standard mp3, 1 for adu formatted mp3
    unsigned int dither_state;
} MPADecodeContext;

/**
 * Context for MP3On4 decoder
 */
typedef struct MP3On4DecodeContext {
    int frames;   ///< number of mp3 frames per block (number of mp3 decoder instances)
    int chan_cfg; ///< channel config number
    MPADecodeContext *mp3decctx[5]; ///< MPADecodeContext for every decoder instance
} MP3On4DecodeContext;

/* layer 3 "granule" */
typedef struct GranuleDef {
    uint8_t scfsi;
    int part2_3_length;
    int big_values;
    int global_gain;
    int scalefac_compress;
    uint8_t block_type;
    uint8_t switch_point;
    int table_select[3];
    int subblock_gain[3];
    uint8_t scalefac_scale;
    uint8_t count1table_select;
    int region_size[3]; /* number of huffman codes in each region */
    int preflag;
    int short_start, long_end; /* long/short band indexes */
    uint8_t scale_factors[40];
    int32_t sb_hybrid[SBLIMIT * 18]; /* 576 samples */
} GranuleDef;

#define MODE_EXT_MS_STEREO 2
#define MODE_EXT_I_STEREO  1

/* layer 3 huffman tables */
typedef struct HuffTable {
    int xsize;
    const uint8_t *bits;
    const uint16_t *codes;
} HuffTable;

#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 huff_quad_vlc[2];
/* computed from band_size_long */
static uint16_t band_index_long[9][23];
/* XXX: free when all decoders are closed */
#define TABLE_4_3_SIZE (8191 + 16)*4
static int8_t  *table_4_3_exp;
static uint32_t *table_4_3_value;
static uint32_t exp_table[512];
static uint32_t expval_table[512][16];
/* intensity stereo coef table */
static int32_t is_table[2][16];
static int32_t is_table_lsf[2][2][16];
static int32_t csa_table[8][4];
static float csa_table_float[8][4];
static int32_t 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(1.0 * (v)), FIXR(0.7937005259 * (v)), FIXR(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 */
};

void ff_mpa_synth_init(MPA_INT *window);
static MPA_INT window[512] __attribute__((aligned(16)));

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

#if defined(USE_HIGHPRECISION) && defined(CONFIG_AUDIO_NONSHORT)
    avctx->sample_fmt= SAMPLE_FMT_S32;
#else
    avctx->sample_fmt= SAMPLE_FMT_S16;
#endif

    if(avctx->antialias_algo != FF_AA_FLOAT)
        s->compute_antialias= compute_antialias_integer;
    else
        s->compute_antialias= compute_antialias_float;

    if (!init && !avctx->parse_only) {
        /* 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_t_C(1) << n) * FRAC_ONE) / ((1 << n) - 1);
            scale_factor_mult[i][0] = MULL(FIXR(1.0 * 2.0), norm);
            scale_factor_mult[i][1] = MULL(FIXR(0.7937005259 * 2.0), norm);
            scale_factor_mult[i][2] = MULL(FIXR(0.6299605249 * 2.0), norm);
            dprintf("%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]);
        }

        ff_mpa_synth_init(window);

        /* huffman decode tables */
        for(i=1;i<16;i++) {
            const HuffTable *h = &mpa_huff_tables[i];
            int xsize, x, y;
            unsigned int n;
            uint8_t tmp_bits [256];
            uint16_t tmp_codes[256];

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

            xsize = h->xsize;
            n = xsize * xsize;

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

            /* XXX: fail test */
            init_vlc(&huff_vlc[i], 7, 256,
                     tmp_bits, 1, 1, tmp_codes, 2, 2, 1);
        }
        for(i=0;i<2;i++) {
            init_vlc(&huff_quad_vlc[i], i == 0 ? 7 : 4, 16,
                     mpa_quad_bits[i], 1, 1, mpa_quad_codes[i], 1, 1, 1);
        }

        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 */
        table_4_3_exp= av_mallocz_static(TABLE_4_3_SIZE * sizeof(table_4_3_exp[0]));
        if(!table_4_3_exp)
            return -1;
        table_4_3_value= av_mallocz_static(TABLE_4_3_SIZE * sizeof(table_4_3_value[0]));
        if(!table_4_3_value)
            return -1;

        int_pow_init();
        for(i=1;i<TABLE_4_3_SIZE;i++) {
            double f, fm;
            int e, m;
            f = pow((double)(i/4), 4.0 / 3.0) * pow(2, (i&3)*0.25);
            fm = frexp(f, &e);
            m = (uint32_t)(fm*(1LL<<31) + 0.5);
            e+= FRAC_BITS - 31 + 5 - 100;

            /* normalized to FRAC_BITS */
            table_4_3_value[i] = m;
//            av_log(NULL, AV_LOG_DEBUG, "%d %d %f\n", i, m, pow((double)i, 4.0 / 3.0));
            table_4_3_exp[i] = -e;
        }
        for(i=0; i<512*16; i++){
            int exponent= (i>>4);
            double f= pow(i&15, 4.0 / 3.0) * pow(2, (exponent-400)*0.25 + FRAC_BITS + 5);
            expval_table[exponent][i&15]= lrintf(f);
            if((i&15)==1)
                exp_table[exponent]= lrintf(f);
        }

        for(i=0;i<7;i++) {
            float f;
            int 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("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;
//            printf("%d %d %d %d\n", FIX(cs), FIX(cs-1), FIX(ca), FIX(cs)-FIX(ca));
//            av_log(NULL, AV_LOG_DEBUG,"%f %f %f %f\n", cs, ca, ca+cs, 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)));
//                av_log(NULL, AV_LOG_DEBUG, "%2d %d %f\n", i,j,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];
            }
        }

#if defined(DEBUG)
        for(j=0;j<8;j++) {
            av_log(avctx, AV_LOG_DEBUG, "win%d=\n", j);
            for(i=0;i<36;i++)
                av_log(avctx, AV_LOG_DEBUG, "%f, ", (double)mdct_win[j][i] / FRAC_ONE);
            av_log(avctx, AV_LOG_DEBUG, "\n");
        }
#endif
        init = 1;
    }

    s->inbuf_index = 0;
    s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
    s->inbuf_ptr = s->inbuf;
#ifdef DEBUG
    s->frame_count = 0;
#endif
    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] = MULH(tmp1<<(s), c);\
}

#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(int32_t *out, int32_t *tab)
{
    int 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 FRAC_BITS <= 15

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

#   if defined(ARCH_POWERPC_405)
        /* signed 16x16 -> 32 multiply add accumulate */
#       define MACS(rt, ra, rb) \
            asm ("maclhw %0, %2, %3" : "=r" (rt) : "0" (rt), "r" (ra), "r" (rb));

        /* signed 16x16 -> 32 multiply */
#       define MULS(ra, rb) \
            ({ int __rt; asm ("mullhw %0, %1, %2" : "=r" (__rt) : "r" (ra), "r" (rb)); __rt; })
#   else
        /* signed 16x16 -> 32 multiply add accumulate */
#       define MACS(rt, ra, rb) rt += (ra) * (rb)

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

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

#   define MULS(ra, rb) MUL64(ra, rb)
#endif

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

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

void ff_mpa_synth_init(MPA_INT *window)
{
    int i;

    /* max = 18760, max sum over all 16 coefs : 44736 */
    for(i=0;i<257;i++) {
        int v;
        v = mpa_enwindow[i];
#if 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 ff_mpa_synth_filter(MPA_INT *synth_buf_ptr, int *synth_buf_offset,
                         MPA_INT *window, int *dither_state,
                         OUT_INT *samples, int incr,
                         int32_t sb_samples[SBLIMIT])
{
    int32_t tmp[32];
    register MPA_INT *synth_buf;
    register const MPA_INT *w, *w2, *p;
    int j, offset, v;
    OUT_INT *samples2;
#if FRAC_BITS <= 15
    int sum, sum2;
#else
    int64_t sum, sum2;
#endif

    dct32(tmp, sb_samples);

    offset = *synth_buf_offset;
    synth_buf = synth_buf_ptr + offset;

    for(j=0;j<32;j++) {
        v = tmp[j];
#if FRAC_BITS <= 15
        /* NOTE: can cause a loss in precision if very high amplitude
           sound */
        if (v > 32767)
            v = 32767;
        else if (v < -32768)
            v = -32768;
#endif
        synth_buf[j] = v;
    }
    /* copy to avoid wrap */
    memcpy(synth_buf + 512, synth_buf, 32 * sizeof(MPA_INT));

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

    sum = *dither_state;
    p = synth_buf + 16;
    SUM8(sum, +=, w, p);
    p = synth_buf + 48;
    SUM8(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, +=, sum2, -=, w, w2, p);
        p = synth_buf + 48 - j;
        SUM8P2(sum, -=, sum2, -=, 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(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 int 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 int 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(int *out, int *in)
{
    int 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];