mpegaudiodec.c

        s0 = t1 + t0;
        s2 = t1 - t0;

        t2 = tmp[i + 1];
        t3 = tmp[i + 3];
        s1 = MULL(t3 + t2, icos36[j]);
        s3 = MULL(t3 - t2, icos36[8 - j]);
        
        t0 = MULL(s0 + s1, icos72[9 + 8 - j]);
        t1 = MULL(s0 - s1, icos72[8 - j]);
        out[18 + 9 + j] = t0;
        out[18 + 8 - j] = t0;
        out[9 + j] = -t1;
        out[8 - j] = t1;
        
        t0 = MULL(s2 + s3, icos72[9+j]);
        t1 = MULL(s2 - s3, icos72[j]);
        out[18 + 9 + (8 - j)] = t0;
        out[18 + j] = t0;
        out[9 + (8 - j)] = -t1;
        out[j] = t1;
        i += 4;
    }

    s0 = tmp[16];
    s1 = MULL(tmp[17], icos36[4]);
    t0 = MULL(s0 + s1, icos72[9 + 4]);
    t1 = MULL(s0 - s1, icos72[4]);
    out[18 + 9 + 4] = t0;
    out[18 + 8 - 4] = t0;
    out[9 + 4] = -t1;
    out[8 - 4] = t1;
}

/* fast header check for resync */
static int check_header(UINT32 header)
{
    /* header */
    if ((header & 0xffe00000) != 0xffe00000)
	return -1;
    /* layer check */
    if (((header >> 17) & 3) == 0)
	return -1;
    /* bit rate */
    if (((header >> 12) & 0xf) == 0xf)
	return -1;
    /* frequency */
    if (((header >> 10) & 3) == 3)
	return -1;
    return 0;
}

/* header + layer + bitrate + freq + lsf/mpeg25 */
#define SAME_HEADER_MASK \
   (0xffe00000 | (3 << 17) | (0xf << 12) | (3 << 10) | (3 << 19))

/* header decoding. MUST check the header before because no
   consistency check is done there. Return 1 if free format found and
   that the frame size must be computed externally */
static int decode_header(MPADecodeContext *s, UINT32 header)
{
    int sample_rate, frame_size, mpeg25, padding;
    int sample_rate_index, bitrate_index;
    if (header & (1<<20)) {
        s->lsf = (header & (1<<19)) ? 0 : 1;
        mpeg25 = 0;
    } else {
        s->lsf = 1;
        mpeg25 = 1;
    }
    
    s->layer = 4 - ((header >> 17) & 3);
    /* extract frequency */
    sample_rate_index = (header >> 10) & 3;
    sample_rate = mpa_freq_tab[sample_rate_index] >> (s->lsf + mpeg25);
    sample_rate_index += 3 * (s->lsf + mpeg25);
    s->sample_rate_index = sample_rate_index;
    s->error_protection = ((header >> 16) & 1) ^ 1;
    s->sample_rate = sample_rate;

    bitrate_index = (header >> 12) & 0xf;
    padding = (header >> 9) & 1;
    //extension = (header >> 8) & 1;
    s->mode = (header >> 6) & 3;
    s->mode_ext = (header >> 4) & 3;
    //copyright = (header >> 3) & 1;
    //original = (header >> 2) & 1;
    //emphasis = header & 3;

    if (s->mode == MPA_MONO)
        s->nb_channels = 1;
    else
        s->nb_channels = 2;
    
    if (bitrate_index != 0) {
        frame_size = mpa_bitrate_tab[s->lsf][s->layer - 1][bitrate_index];
        s->bit_rate = frame_size * 1000;
        switch(s->layer) {
        case 1:
            frame_size = (frame_size * 12000) / sample_rate;
            frame_size = (frame_size + padding) * 4;
            break;
        case 2:
            frame_size = (frame_size * 144000) / sample_rate;
            frame_size += padding;
            break;
        default:
        case 3:
            frame_size = (frame_size * 144000) / (sample_rate << s->lsf);
            frame_size += padding;
            break;
        }
        s->frame_size = frame_size;
    } else {
        /* if no frame size computed, signal it */
        if (!s->free_format_frame_size)
            return 1;
        /* free format: compute bitrate and real frame size from the
           frame size we extracted by reading the bitstream */
        s->frame_size = s->free_format_frame_size;
        switch(s->layer) {
        case 1:
            s->frame_size += padding  * 4;
            s->bit_rate = (s->frame_size * sample_rate) / 48000;
            break;
        case 2:
            s->frame_size += padding;
            s->bit_rate = (s->frame_size * sample_rate) / 144000;
            break;
        default:
        case 3:
            s->frame_size += padding;
            s->bit_rate = (s->frame_size * (sample_rate << s->lsf)) / 144000;
            break;
        }
    }
    
#if defined(DEBUG)
    printf("layer%d, %d Hz, %d kbits/s, ",
           s->layer, s->sample_rate, s->bit_rate);
    if (s->nb_channels == 2) {
        if (s->layer == 3) {
            if (s->mode_ext & MODE_EXT_MS_STEREO)
                printf("ms-");
            if (s->mode_ext & MODE_EXT_I_STEREO)
                printf("i-");
        }
        printf("stereo");
    } else {
        printf("mono");
    }
    printf("\n");
#endif
    return 0;
}

/* return the number of decoded frames */
static int mp_decode_layer1(MPADecodeContext *s)
{
    int bound, i, v, n, ch, j, mant;
    UINT8 allocation[MPA_MAX_CHANNELS][SBLIMIT];
    UINT8 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;
}

/* bitrate is in kb/s */
int l2_select_table(int bitrate, int nb_channels, int freq, int lsf)
{
    int ch_bitrate, table;
    
    ch_bitrate = bitrate / nb_channels;
    if (!lsf) {
        if ((freq == 48000 && ch_bitrate >= 56) ||
            (ch_bitrate >= 56 && ch_bitrate <= 80)) 
            table = 0;
        else if (freq != 48000 && ch_bitrate >= 96) 
            table = 1;
        else if (freq != 32000 && ch_bitrate <= 48) 
            table = 2;
        else 
            table = 3;
    } else {
        table = 4;
    }
    return table;
}

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 = l2_select_table(s->bit_rate / 1000, s->nb_channels, 
                            s->sample_rate, s->lsf);
    sblimit = sblimit_table[table];
    alloc_table = alloc_tables[table];

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

    dprintf("bound=%d sblimit=%d\n", 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;
    }

#ifdef DEBUG
    {
        for(ch=0;ch<s->nb_channels;ch++) {
            for(i=0;i<sblimit;i++)
                printf(" %d", bit_alloc[ch][i]);
            printf("\n");
        }
    }
#endif

    /* 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;
                }
            }
        }
    }

#ifdef DEBUG
    for(ch=0;ch<s->nb_channels;ch++) {
        for(i=0;i<sblimit;i++) {
            if (bit_alloc[ch][i]) {
                sf = scale_factors[ch][i];
                printf(" %d %d %d", sf[0], sf[1], sf[2]);
            } else {
                printf(" -");
            }
        }
        printf("\n");
    }
#endif

    /* 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 = quant_bits[qindex];
                        if (bits < 0) {
                            /* 3 values at the same time */
                            v = get_bits(&s->gb, -bits);
                            steps = 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 = quant_bits[qindex];
                    if (bits < 0) {
                        /* 3 values at the same time */
                        v = get_bits(&s->gb, -bits);
                        steps = 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;
}

/*
 * Seek back in the stream for backstep bytes (at most 511 bytes)
 */
static void seek_to_maindata(MPADecodeContext *s, long backstep)
{
    UINT8 *ptr;

    /* compute current position in stream */
#ifdef ALT_BITSTREAM_READER
    ptr = s->gb.buffer + (s->gb.index>>3);
#else
    ptr = s->gb.buf_ptr - (s->gb.bit_cnt >> 3);
#endif    
    /* copy old data before current one */
    ptr -= backstep;
    memcpy(ptr, s->inbuf1[s->inbuf_index ^ 1] + 
           BACKSTEP_SIZE + s->old_frame_size - backstep, backstep);
    /* init get bits again */
    init_get_bits(&s->gb, ptr, s->frame_size + backstep);

    /* prepare next buffer */
    s->inbuf_index ^= 1;
    s->inbuf = &s->inbuf1[s->inbuf_index][BACKSTEP_SIZE];
    s->old_frame_size = s->frame_size;
}

static inline void lsf_sf_expand(int *slen,
                                 int sf, int n1, int n2, int n3)
{
    if (n3) {
        slen[3] = sf % n3;
        sf /= n3;
    } else {
        slen[3] = 0;
    }
    if (n2) {
        slen[2] = sf % n2;
        sf /= n2;
    } else {
        slen[2] = 0;
    }
    slen[1] = sf % n1;
    sf /= n1;
    slen[0] = sf;
}

static void exponents_from_scale_factors(MPADecodeContext *s, 
                                         GranuleDef *g,
                                         INT16 *exponents)
{
    const UINT8 *bstab, *pretab;
    int len, i, j, k, l, v0, shift, gain, gains[3];
    INT16 *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);
        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);
                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 int huffman_decode(MPADecodeContext *s, GranuleDef *g,
                          INT16 *exponents, int end_pos)
{
    int s_index;
    int linbits, code, x, y, l, v, i, j, k, pos;
    UINT8 *last_buf_ptr;
    UINT32 last_bit_buf;
    int last_bit_cnt;
    VLC *vlc;
    UINT8 *code_table;

    /* low frequencies (called big values) */
    s_index = 0;
    for(i=0;i<3;i++) {
        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];
        code_table = huff_code_table[l];

        /* read huffcode and compute each couple */
        for(;j>0;j--) {
            if (get_bits_count(&s->gb) >= end_pos)
                break;
            if (code_table) {
                code = get_vlc(&s->gb, vlc);
                if (code < 0)
                    return -1;
                y = code_table[code];
                x = y >> 4;
                y = y & 0x0f;
            } else {
                x = 0;
                y = 0;
            }
            dprintf("region=%d n=%d x=%d y=%d exp=%d\n", 
                    i, g->region_size[i] - j, x, y, exponents[s_index]);
            if (x) {
                if (x == 15)
                    x += get_bitsz(&s->gb, linbits);
                v = l3_unscale(x, exponents[s_index]);
                if (get_bits1(&s->gb))
                    v = -v;
            } else {
                v = 0;
            }
            g->sb_hybrid[s_index++] = v;
            if (y) {
                if (y == 15)
                    y += get_bitsz(&s->gb, linbits);
                v = l3_unscale(y, exponents[s_index]);
                if (get_bits1(&s->gb))
                    v = -v;
            } else {
                v = 0;
            }
            g->sb_hybrid[s_index++] = v;
        }
    }
            
    /* high frequencies */
    vlc = &huff_quad_vlc[g->count1table_select];
    last_buf_ptr = NULL;
    last_bit_buf = 0;
    last_bit_cnt = 0;
    while (s_index <= 572) {
        pos = get_bits_count(&s->gb);
        if (pos >= end_pos) {
            if (pos > end_pos && last_buf_ptr != NULL) {
                /* some encoders generate an incorrect size for this
                   part. We must go back into the data */
                s_index -= 4;
#ifdef ALT_BITSTREAM_READER
                s->gb.buffer = last_buf_ptr;
                s->gb.index = last_bit_cnt;
#else
                s->gb.buf_ptr = last_buf_ptr;
                s->gb.bit_buf = last_bit_buf;
                s->gb.bit_cnt = last_bit_cnt;
#endif            
            }
            break;
        }
#ifdef ALT_BITSTREAM_READER
        last_buf_ptr = s->gb.buffer;
        last_bit_cnt = s->gb.index;
#else
        last_buf_ptr = s->gb.buf_ptr;
        last_bit_buf = s->gb.bit_buf;
        last_bit_cnt = s->gb.bit_cnt;
#endif
        
        code = get_vlc(&s->gb, vlc);
        dprintf("t=%d code=%d\n", g->count1table_select, code);
        if (code < 0)
            return -1;
        for(i=0;i<4;i++) {
            if (code & (8 >> i)) {
                /* non zero value. Could use a hand coded function for
                   'one' value */
                v = l3_unscale(1, exponents[s_index]);
                if(get_bits1(&s->gb))
                    v = -v;
            } else {
                v = 0;
            }
            g->sb_hybrid[s_index++] = v;
        }
    }
    while (s_index < 576)
        g->sb_hybrid[s_index++] = 0;
    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, k, len;
    INT32 *ptr, *dst, *ptr1;
    INT32 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;
        for(k=0;k<3;k++) {
            dst = tmp + k;
            for(j=len;j>0;j--) {
                *dst = *ptr++;
                dst += 3;
            }
        }
        memcpy(ptr1, tmp, len * 3 * sizeof(INT32));
    }
}

#define ISQRT2 FIXR(0.70710678118654752440)

static void compute_stereo(MPADecodeContext *s,
                           GranuleDef *g0, GranuleDef *g1)
{
    int i, j, k, l;
    INT32 v1, v2;
    int sf_max, tmp0, tmp1, sf, len, non_zero_found;
    INT32 (*is_tab)[16];
    INT32 *tab0, *tab1;
    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] = MULL(tmp0, v1);
                        tab1[j] = MULL(tmp0, v2);
                    }
                } 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] = MULL(tmp0 + tmp1, ISQRT2);
                            tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
                        }
                    }
                }
            }
        }

        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] = MULL(tmp0, v1);
                    tab1[j] = MULL(tmp0, v2);
                }
            } 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] = MULL(tmp0 + tmp1, ISQRT2);
                        tab1[j] = MULL(tmp0 - tmp1, ISQRT2);
                    }
                }
            }
        }
    } 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(MPADecodeContext *s,
                              GranuleDef *g)
{
    INT32 *ptr, *p0, *p1, *csa;
    int n, tmp0, tmp1, i, j;

    /* 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--) {
        p0 = ptr - 1;
        p1 = ptr;
        csa = &csa_table[0][0];
        for(j=0;j<8;j++) {
            tmp0 = *p0;
            tmp1 = *p1;
            *p0 = FRAC_RND(MUL64(tmp0, csa[0]) - MUL64(tmp1, csa[1]));
            *p1 = FRAC_RND(MUL64(tmp0, csa[1]) + MUL64(tmp1, csa[0]));
            p0--;
            p1++;
            csa += 2;
        }
        ptr += 18;
    }
}

static void compute_imdct(MPADecodeContext *s,
                          GranuleDef *g, 
                          INT32 *sb_samples,
                          INT32 *mdct_buf)
{
    INT32 *ptr, *win, *win1, *buf, *buf2, *out_ptr, *ptr1;
    INT32 in[6];
    INT32 out[36];
    INT32 out2[12];
    int i, j, k, mdct_long_end, v, sblimit;

    /* find last non zero block */
    ptr = g->sb_hybrid + 576;
    ptr1 = g->sb_hybrid + 2 * 18;
    while (ptr >= ptr1) {
        ptr -= 6;
        v = ptr[0] | ptr[1] | ptr[2] | ptr[3] | ptr[4] | ptr[5];
        if (v != 0)
            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++) {
        imdct36(out, ptr);
        /* 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));
        for(i=0;i<18;i++) {
            *out_ptr = MULL(out[i], win[i]) + buf[i];
            buf[i] = MULL(out[i + 18], win[i + 18]);
            out_ptr += SBLIMIT;
        }
        ptr += 18;
        buf += 18;
    }
    for(j=mdct_long_end;j<sblimit;j++) {
        for(i=0;i<6;i++) {
            out[i] = 0;
            out[6 + i] = 0;
            out[30+i] = 0;
        }
        /* select frequency inversion */
        win = mdct_win[2] + ((4 * 36) & -(j & 1));
        buf2 = out + 6;
        for(k=0;k<3;k++) {
            /* reorder input for short mdct */
            ptr1 = ptr + k;
            for(i=0;i<6;i++) {
                in[i] = *ptr1;
                ptr1 += 3;
            }
            imdct12(out2, in);
            /* apply 12 point window and do small overlap */
            for(i=0;i<6;i++) {
                buf2[i] = MULL(out2[i], win[i]) + buf2[i];
                buf2[i + 6] = MULL(out2[i + 6], win[i + 6]);
            }
            buf2 += 6;
        }
        /* overlap */
        out_ptr = sb_samples + j;
        for(i=0;i<18;i++) {
            *out_ptr = out[i] + buf[i];
            buf[i] = out[i + 18];
            out_ptr += SBLIMIT;
        }
        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;
    }
}

#if defined(DEBUG)
void sample_dump(int fnum, INT32 *tab, int n)
{
    static FILE *files[16], *f;
    char buf[512];
    int i;
    INT32 v;
    
    f = files[fnum];
    if (!f) {
        sprintf(buf, "/tmp/out%d.%s.pcm", 
                fnum, 
#ifdef USE_HIGHPRECISION
                "hp"
#else
                "lp"
#endif
                );
        f = fopen(buf, "w");
        if (!f)
            return;
        files[fnum] = f;
    }
    
    if (fnum == 0) {
        static int pos = 0;
        printf("pos=%d\n", pos);
        for(i=0;i<n;i++) {
            printf(" %0.4f", (double)tab[i] / FRAC_ONE);
            if ((i % 18) == 17)
                printf("\n");
        }
        pos += n;
    }
    for(i=0;i<n;i++) {
        /* normalize to 23 frac bits */
        v = tab[i] << (23 - FRAC_BITS);
        fwrite(&v, 1, sizeof(INT32), f);
    }