mpegaudiodec.c

        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);
    }
}

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;
            INTFLOAT 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){
                    v = RENAME(expval_table)[ exponent ][ x ];
//                      v = RENAME(expval_table)[ (exponent&3) ][ x ] >> FFMIN(0 - (exponent>>2), 31);
                }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){
                    v = 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){
                    v = 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};
            INTFLOAT v;
            int pos= s_index+idxtab[code];
            code ^= 8>>idxtab[code];
            v = RENAME(exp_table)[ exponents[pos] ];
//            v = RENAME(exp_table)[ (exponents[pos]&3) ] >> FFMIN(0 - (exponents[pos]>>2), 31);
            if(get_bits1(&s->gb)) //FIXME try to flip the sign bit in int32_t, same above
                v = -v;
            g->sb_hybrid[pos] = v;
        }
        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));
  }