vorbis.c

    if (vorbis_parse_id_hdr(vc)) {
        av_log(avccontext, AV_LOG_ERROR, "Id header corrupt.\n");
        vorbis_free(vc);
        return -1;
    }

    init_get_bits(gb, header_start[2], header_len[2]*8);
    hdr_type=get_bits(gb, 8);
    if (hdr_type!=5) {
        av_log(avccontext, AV_LOG_ERROR, "Third header is not the setup header.\n");
        return -1;
    }
    if (vorbis_parse_setup_hdr(vc)) {
        av_log(avccontext, AV_LOG_ERROR, "Setup header corrupt.\n");
        vorbis_free(vc);
        return -1;
    }

    avccontext->channels = vc->audio_channels;
    avccontext->sample_rate = vc->audio_samplerate;

    return 0 ;
}

// Decode audiopackets -------------------------------------------------

// Read and decode floor

static uint_fast8_t vorbis_floor0_decode(vorbis_context *vc,
                                         vorbis_floor_data *vfu, float *vec) {
    vorbis_floor0 * vf=&vfu->t0;
    float * lsp=vf->lsp;
    uint_fast32_t amplitude;
    uint_fast32_t book_idx;
    uint_fast8_t blockflag=vc->modes[vc->mode_number].blockflag;

    amplitude=get_bits(&vc->gb, vf->amplitude_bits);
    if (amplitude>0) {
        float last = 0;
        uint_fast16_t lsp_len = 0;
        uint_fast16_t idx;
        vorbis_codebook codebook;

        book_idx=get_bits(&vc->gb, ilog(vf->num_books));
        if ( book_idx >= vf->num_books ) {
            av_log( vc->avccontext, AV_LOG_ERROR,
                    "floor0 dec: booknumber too high!\n" );
            //FIXME: look above
        }
        AV_DEBUG( "floor0 dec: booknumber: %u\n", book_idx );
        codebook=vc->codebooks[vf->book_list[book_idx]];

        while (lsp_len<vf->order) {
            int vec_off;

            AV_DEBUG( "floor0 dec: book dimension: %d\n", codebook.dimensions );
            AV_DEBUG( "floor0 dec: maximum depth: %d\n", codebook.maxdepth );
            /* read temp vector */
            vec_off=get_vlc2(&vc->gb,
                             codebook.vlc.table,
                             codebook.nb_bits,
                             codebook.maxdepth ) *
                             codebook.dimensions;
            AV_DEBUG( "floor0 dec: vector offset: %d\n", vec_off );
            /* copy each vector component and add last to it */
            for (idx=0; idx<codebook.dimensions; ++idx) {
                lsp[lsp_len+idx]=codebook.codevectors[vec_off+idx]+last;
            }
            last=lsp[lsp_len+idx-1]; /* set last to last vector component */

            lsp_len += codebook.dimensions;
        }
#ifdef V_DEBUG
        /* DEBUG: output lsp coeffs */
        {
            int idx;
            for ( idx = 0; idx < lsp_len; ++idx )
                AV_DEBUG("floor0 dec: coeff at %d is %f\n", idx, lsp[idx] );
        }
#endif

        /* synthesize floor output vector */
        {
            int i;
            int order=vf->order;
            float wstep=M_PI/vf->bark_map_size;

            for(i=0;i<order;i++) { lsp[i]=2.0f*cos(lsp[i]); }

            AV_DEBUG("floor0 synth: map_size=%d; m=%d; wstep=%f\n",
                     vf->map_size, order, wstep);

            i=0;
            while(i<vf->map_size[blockflag]) {
                int j, iter_cond=vf->map[blockflag][i];
                float p=0.5f;
                float q=0.5f;
                float two_cos_w=2.0f*cos(wstep*iter_cond); // needed all times

                /* similar part for the q and p products */
                for(j=0;j<order;j+=2) {
                    q *= lsp[j]  -two_cos_w;
                    p *= lsp[j+1]-two_cos_w;
                }
                if(j==order) { // even order
                    p *= p*(2.0f-two_cos_w);
                    q *= q*(2.0f+two_cos_w);
                }
                else { // odd order
                    q *= two_cos_w-lsp[j]; // one more time for q

                    /* final step and square */
                    p *= p*(4.f-two_cos_w*two_cos_w);
                    q *= q;
                }

                /* calculate linear floor value */
                {
                    q=exp( (
                             ( (amplitude*vf->amplitude_offset)/
                               (((1<<vf->amplitude_bits)-1) * sqrt(p+q)) )
                             - vf->amplitude_offset ) * .11512925f
                         );
                }

                /* fill vector */
                do { vec[i]=q; ++i; }while(vf->map[blockflag][i]==iter_cond);
            }
        }
    }
    else {
        /* this channel is unused */
        return 1;
    }

    AV_DEBUG(" Floor0 decoded\n");

    return 0;
}
static uint_fast8_t vorbis_floor1_decode(vorbis_context *vc, vorbis_floor_data *vfu, float *vec) {
    vorbis_floor1 * vf=&vfu->t1;
    GetBitContext *gb=&vc->gb;
    uint_fast16_t range_v[4]={ 256, 128, 86, 64 };
    uint_fast16_t range=range_v[vf->multiplier-1];
    uint_fast16_t floor1_Y[vf->x_list_dim];
    uint_fast16_t floor1_Y_final[vf->x_list_dim];
    uint_fast8_t floor1_flag[vf->x_list_dim];
    uint_fast8_t class_;
    uint_fast8_t cdim;
    uint_fast8_t cbits;
    uint_fast8_t csub;
    uint_fast8_t cval;
    int_fast16_t book;
    uint_fast16_t offset;
    uint_fast16_t i,j;
    uint_fast16_t *floor_x_sort=vf->x_list_order;
    /*u*/int_fast16_t adx, ady, off, predicted; // WTF ? dy/adx= (unsigned)dy/adx ?
    int_fast16_t dy, err;
    uint_fast16_t lx,hx, ly, hy=0;


    if (!get_bits1(gb)) return 1; // silence

// Read values (or differences) for the floor's points

    floor1_Y[0]=get_bits(gb, ilog(range-1));
    floor1_Y[1]=get_bits(gb, ilog(range-1));

    AV_DEBUG("floor 0 Y %d floor 1 Y %d \n", floor1_Y[0], floor1_Y[1]);

    offset=2;
    for(i=0;i<vf->partitions;++i) {
        class_=vf->partition_class[i];
        cdim=vf->class_dimensions[class_];
        cbits=vf->class_subclasses[class_];
        csub=(1<<cbits)-1;
        cval=0;

        AV_DEBUG("Cbits %d \n", cbits);

        if (cbits) { // this reads all subclasses for this partition's class
            cval=get_vlc2(gb, vc->codebooks[vf->class_masterbook[class_]].vlc.table,
            vc->codebooks[vf->class_masterbook[class_]].nb_bits, 3);
        }

        for(j=0;j<cdim;++j) {
            book=vf->subclass_books[class_][cval & csub];

            AV_DEBUG("book %d Cbits %d cval %d  bits:%d \n", book, cbits, cval, get_bits_count(gb));

            cval=cval>>cbits;
            if (book>0) {
                floor1_Y[offset+j]=get_vlc2(gb, vc->codebooks[book].vlc.table,
                vc->codebooks[book].nb_bits, 3);
            } else {
                floor1_Y[offset+j]=0;
            }

            AV_DEBUG(" floor(%d) = %d \n", vf->x_list[offset+j], floor1_Y[offset+j]);
        }
        offset+=cdim;
    }

// Amplitude calculation from the differences

    floor1_flag[0]=1;
    floor1_flag[1]=1;
    floor1_Y_final[0]=floor1_Y[0];
    floor1_Y_final[1]=floor1_Y[1];

    for(i=2;i<vf->x_list_dim;++i) {
        uint_fast16_t val, highroom, lowroom, room;
        uint_fast16_t high_neigh_offs;
        uint_fast16_t low_neigh_offs;

        low_neigh_offs=vf->low_neighbour[i];
        high_neigh_offs=vf->high_neighbour[i];
        dy=floor1_Y_final[high_neigh_offs]-floor1_Y_final[low_neigh_offs];  // render_point begin
        adx=vf->x_list[high_neigh_offs]-vf->x_list[low_neigh_offs];
        ady= ABS(dy);
        err=ady*(vf->x_list[i]-vf->x_list[low_neigh_offs]);
        off=(int16_t)err/(int16_t)adx;
        if (dy<0) {
            predicted=floor1_Y_final[low_neigh_offs]-off;
        } else {
            predicted=floor1_Y_final[low_neigh_offs]+off;
        } // render_point end

        val=floor1_Y[i];
        highroom=range-predicted;
        lowroom=predicted;
        if (highroom < lowroom) {
            room=highroom*2;
        } else {
            room=lowroom*2;   // SPEC mispelling
        }
        if (val) {
            floor1_flag[low_neigh_offs]=1;
            floor1_flag[high_neigh_offs]=1;
            floor1_flag[i]=1;
            if (val>=room) {
                if (highroom > lowroom) {
                    floor1_Y_final[i]=val-lowroom+predicted;
                } else {
                    floor1_Y_final[i]=predicted-val+highroom-1;
                }
            } else {
                if (val & 1) {
                    floor1_Y_final[i]=predicted-(val+1)/2;
                } else {
                    floor1_Y_final[i]=predicted+val/2;
                }
            }
        } else {
            floor1_flag[i]=0;
            floor1_Y_final[i]=predicted;
        }

        AV_DEBUG(" Decoded floor(%d) = %d / val %d \n", vf->x_list[i], floor1_Y_final[i], val);
    }

// Curve synth - connect the calculated dots and convert from dB scale FIXME optimize ?

    hx=0;
    lx=0;
    ly=floor1_Y_final[0]*vf->multiplier;  // conforms to SPEC

    vec[0]=floor1_inverse_db_table[ly];

    for(i=1;i<vf->x_list_dim;++i) {
        AV_DEBUG(" Looking at post %d \n", i);

        if (floor1_flag[floor_x_sort[i]]) {   // SPEC mispelled
            int_fast16_t x, y, dy, base, sy; // if uncommented: dy = -32 adx = 2  base = 2blablabla ?????

            hy=floor1_Y_final[floor_x_sort[i]]*vf->multiplier;
            hx=vf->x_list[floor_x_sort[i]];

            dy=hy-ly;
            adx=hx-lx;
            ady= (dy<0) ? -dy:dy;//ABS(dy);
            base=(int16_t)dy/(int16_t)adx;

            AV_DEBUG(" dy %d  adx %d base %d = %d \n", dy, adx, base, dy/adx);

            x=lx;
            y=ly;
            err=0;
            if (dy<0) {
                sy=base-1;
            } else {
                sy=base+1;
            }
            ady=ady-(base<0 ? -base : base)*adx;
            vec[x]=floor1_inverse_db_table[y];

            AV_DEBUG(" vec[ %d ] = %d \n", x, y);

            for(x=lx+1;(x<hx) && (x<vf->x_list[1]);++x) {
                err+=ady;
                if (err>=adx) {
                    err-=adx;
                    y+=sy;
                } else {
                    y+=base;
                }
                vec[x]=floor1_inverse_db_table[y];

                AV_DEBUG(" vec[ %d ] = %d \n", x, y);
            }

/*            for(j=1;j<hx-lx+1;++j) {  // iterating render_point
                dy=hy-ly;
                adx=hx-lx;
                ady= dy<0 ? -dy : dy;
                err=ady*j;
                off=err/adx;
                if (dy<0) {
                    predicted=ly-off;
                } else {
                    predicted=ly+off;
                }
                if (lx+j < vf->x_list[1]) {
                    vec[lx+j]=floor1_inverse_db_table[predicted];
                }
            }*/

            lx=hx;
            ly=hy;
        }
    }

    if (hx<vf->x_list[1]) {
        for(i=hx;i<vf->x_list[1];++i) {
            vec[i]=floor1_inverse_db_table[hy];
        }
    }

    AV_DEBUG(" Floor decoded\n");

    return 0;
}

// Read and decode residue

static int vorbis_residue_decode(vorbis_context *vc, vorbis_residue *vr, uint_fast8_t ch, uint_fast8_t *do_not_decode, float *vec, uint_fast16_t vlen) {
    GetBitContext *gb=&vc->gb;
    uint_fast8_t c_p_c=vc->codebooks[vr->classbook].dimensions;
    uint_fast16_t n_to_read=vr->end-vr->begin;
    uint_fast16_t ptns_to_read=n_to_read/vr->partition_size;
    uint_fast8_t classifs[ptns_to_read*vc->audio_channels];
    uint_fast8_t pass;
    uint_fast8_t ch_used;
    uint_fast8_t i,j,l;
    uint_fast16_t k;

    if (vr->type==2) {
        for(j=1;j<ch;++j) {
                do_not_decode[0]&=do_not_decode[j];  // FIXME - clobbering input
        }
        if (do_not_decode[0]) return 0;
        ch_used=1;
    } else {
        ch_used=ch;
    }

    AV_DEBUG(" residue type 0/1/2 decode begin, ch: %d  cpc %d  \n", ch, c_p_c);

    for(pass=0;pass<=vr->maxpass;++pass) { // FIXME OPTIMIZE?
        uint_fast16_t voffset;
        uint_fast16_t partition_count;
        uint_fast16_t j_times_ptns_to_read;

        voffset=vr->begin;
        for(partition_count=0;partition_count<ptns_to_read;) {  // SPEC        error
            if (!pass) {
                for(j_times_ptns_to_read=0, j=0;j<ch_used;++j) {
                    if (!do_not_decode[j]) {
                        uint_fast32_t temp=get_vlc2(gb, vc->codebooks[vr->classbook].vlc.table,
                        vc->codebooks[vr->classbook].nb_bits, 3);

                        AV_DEBUG("Classword: %d \n", temp);

                        assert(vr->classifications > 1 && temp<=65536); //needed for inverse[]
                        for(i=0;i<c_p_c;++i) {
                            uint_fast32_t temp2;

                            temp2=(((uint_fast64_t)temp) * inverse[vr->classifications])>>32;
                            if (partition_count+c_p_c-1-i < ptns_to_read) {
                                classifs[j_times_ptns_to_read+partition_count+c_p_c-1-i]=temp-temp2*vr->classifications;
                            }
                            temp=temp2;
                        }
                    }
                    j_times_ptns_to_read+=ptns_to_read;
                }
            }
            for(i=0;(i<c_p_c) && (partition_count<ptns_to_read);++i) {
                for(j_times_ptns_to_read=0, j=0;j<ch_used;++j) {
                    uint_fast16_t voffs;

                    if (!do_not_decode[j]) {
                        uint_fast8_t vqclass=classifs[j_times_ptns_to_read+partition_count];
                        int_fast16_t vqbook=vr->books[vqclass][pass];

                        if (vqbook>=0) {
                            uint_fast16_t coffs;
                            uint_fast8_t dim= vc->codebooks[vqbook].dimensions;
                            uint_fast16_t step= dim==1 ? vr->partition_size
                                              : FASTDIV(vr->partition_size, dim);
                            vorbis_codebook codebook= vc->codebooks[vqbook];

                            if (vr->type==0) {

                                voffs=voffset+j*vlen;
                                for(k=0;k<step;++k) {
                                    coffs=get_vlc2(gb, codebook.vlc.table, codebook.nb_bits, 3) * codebook.dimensions;
                                    for(l=0;l<codebook.dimensions;++l) {
                                        vec[voffs+k+l*step]+=codebook.codevectors[coffs+l];  // FPMATH
                                    }
                                }
                            }
                            else if (vr->type==1) {
                                voffs=voffset+j*vlen;
                                for(k=0;k<step;++k) {
                                    coffs=get_vlc2(gb, codebook.vlc.table, codebook.nb_bits, 3) * codebook.dimensions;
                                    for(l=0;l<codebook.dimensions;++l, ++voffs) {
                                        vec[voffs]+=codebook.codevectors[coffs+l];  // FPMATH

                                        AV_DEBUG(" pass %d offs: %d curr: %f change: %f cv offs.: %d  \n", pass, voffs, vec[voffs], codebook.codevectors[coffs+l], coffs);
                                    }
                                }
                            }
                            else if (vr->type==2 && ch==2 && (voffset&1)==0 && (codebook.dimensions&1)==0) { // most frequent case optimized
                                voffs=voffset>>1;

                                for(k=0;k<step;++k) {
                                    coffs=get_vlc2(gb, codebook.vlc.table, codebook.nb_bits, 3) * codebook.dimensions;
                                    for(l=0;l<codebook.dimensions;l+=2, voffs++) {
                                        vec[voffs     ]+=codebook.codevectors[coffs+l  ];  // FPMATH
                                        vec[voffs+vlen]+=codebook.codevectors[coffs+l+1];  // FPMATH

                                        AV_DEBUG(" pass %d offs: %d curr: %f change: %f cv offs.: %d+%d  \n", pass, voffset/ch+(voffs%ch)*vlen, vec[voffset/ch+(voffs%ch)*vlen], codebook.codevectors[coffs+l], coffs, l);
                                    }
                                }

                            }
                            else if (vr->type==2) {
                                voffs=voffset;

                                for(k=0;k<step;++k) {
                                    coffs=get_vlc2(gb, codebook.vlc.table, codebook.nb_bits, 3) * codebook.dimensions;
                                    for(l=0;l<codebook.dimensions;++l, ++voffs) {
                                        vec[voffs/ch+(voffs%ch)*vlen]+=codebook.codevectors[coffs+l];  // FPMATH FIXME use if and counter instead of / and %

                                        AV_DEBUG(" pass %d offs: %d curr: %f change: %f cv offs.: %d+%d  \n", pass, voffset/ch+(voffs%ch)*vlen, vec[voffset/ch+(voffs%ch)*vlen], codebook.codevectors[coffs+l], coffs, l);
                                    }
                                }
                            } else {
                                av_log(vc->avccontext, AV_LOG_ERROR, " Invalid residue type while residue decode?! \n");
                                return 1;
                            }
                        }
                    }
                    j_times_ptns_to_read+=ptns_to_read;
                }
                ++partition_count;
                voffset+=vr->partition_size;
            }
        }
    }
    return 0;
}

void vorbis_inverse_coupling(float *mag, float *ang, int blocksize)
{
    int i;
    for(i=0; i<blocksize; i++)
    {
        if (mag[i]>0.0) {
            if (ang[i]>0.0) {
                ang[i]=mag[i]-ang[i];
            } else {
                float temp=ang[i];
                ang[i]=mag[i];
                mag[i]+=temp;
            }
        } else {
            if (ang[i]>0.0) {
                ang[i]+=mag[i];
            } else {
                float temp=ang[i];
                ang[i]=mag[i];
                mag[i]-=temp;
            }
        }
    }
}

// Decode the audio packet using the functions above

static int vorbis_parse_audio_packet(vorbis_context *vc) {
    GetBitContext *gb=&vc->gb;

    uint_fast8_t previous_window=0,next_window=0;
    uint_fast8_t mode_number;
    uint_fast16_t blocksize;
    int_fast32_t i,j;
    uint_fast8_t no_residue[vc->audio_channels];
    uint_fast8_t do_not_decode[vc->audio_channels];
    vorbis_mapping *mapping;
    float *ch_res_ptr=vc->channel_residues;
    float *ch_floor_ptr=vc->channel_floors;
    uint_fast8_t res_chan[vc->audio_channels];
    uint_fast8_t res_num=0;
    int_fast16_t retlen=0;
    uint_fast16_t saved_start=0;
    float fadd_bias = vc->add_bias;

    if (get_bits1(gb)) {
        av_log(vc->avccontext, AV_LOG_ERROR, "Not a Vorbis I audio packet.\n");
        return -1; // packet type not audio
    }

    if (vc->mode_count==1) {
        mode_number=0;
    } else {
        mode_number=get_bits(gb, ilog(vc->mode_count-1));
    }
    vc->mode_number=mode_number;
    mapping=&vc->mappings[vc->modes[mode_number].mapping];

    AV_DEBUG(" Mode number: %d , mapping: %d , blocktype %d \n", mode_number, vc->modes[mode_number].mapping, vc->modes[mode_number].blockflag);

    if (vc->modes[mode_number].blockflag) {
        previous_window=get_bits1(gb);
        next_window=get_bits1(gb);
    }

    blocksize=vc->modes[mode_number].blockflag ? vc->blocksize_1 : vc->blocksize_0;
    memset(ch_res_ptr, 0, sizeof(float)*vc->audio_channels*blocksize/2); //FIXME can this be removed ?
    memset(ch_floor_ptr, 0, sizeof(float)*vc->audio_channels*blocksize/2); //FIXME can this be removed ?

// Decode floor

    for(i=0;i<vc->audio_channels;++i) {
        vorbis_floor *floor;
        if (mapping->submaps>1) {
            floor=&vc->floors[mapping->submap_floor[mapping->mux[i]]];
        } else {
            floor=&vc->floors[mapping->submap_floor[0]];
        }

        no_residue[i]=floor->decode(vc, &floor->data, ch_floor_ptr);
        ch_floor_ptr+=blocksize/2;
    }

// Nonzero vector propagate

    for(i=mapping->coupling_steps-1;i>=0;--i) {
        if (!(no_residue[mapping->magnitude[i]] & no_residue[mapping->angle[i]])) {
            no_residue[mapping->magnitude[i]]=0;
            no_residue[mapping->angle[i]]=0;
        }
    }

// Decode residue

    for(i=0;i<mapping->submaps;++i) {
        vorbis_residue *residue;
        uint_fast8_t ch=0;

        for(j=0;j<vc->audio_channels;++j) {
            if ((mapping->submaps==1) || (i=mapping->mux[j])) {
                res_chan[j]=res_num;
                if (no_residue[j]) {
                    do_not_decode[ch]=1;
                } else {
                    do_not_decode[ch]=0;
                }
                ++ch;
                ++res_num;
            }
        }
        residue=&vc->residues[mapping->submap_residue[i]];
        vorbis_residue_decode(vc, residue, ch, do_not_decode, ch_res_ptr, blocksize/2);

        ch_res_ptr+=ch*blocksize/2;
    }

// Inverse coupling

    for(i=mapping->coupling_steps-1;i>=0;--i) { //warning: i has to be signed
        float *mag, *ang;

        mag=vc->channel_residues+res_chan[mapping->magnitude[i]]*blocksize/2;
        ang=vc->channel_residues+res_chan[mapping->angle[i]]*blocksize/2;
        vc->dsp.vorbis_inverse_coupling(mag, ang, blocksize/2);
    }

// Dotproduct

    for(j=0, ch_floor_ptr=vc->channel_floors;j<vc->audio_channels;++j,ch_floor_ptr+=blocksize/2) {
        ch_res_ptr=vc->channel_residues+res_chan[j]*blocksize/2;
        vc->dsp.vector_fmul(ch_floor_ptr, ch_res_ptr, blocksize/2);
    }

// MDCT, overlap/add, save data for next overlapping  FPMATH

    for(j=0;j<vc->audio_channels;++j) {
        uint_fast8_t step=vc->audio_channels;
        uint_fast16_t k;
        float *saved=vc->saved+j*vc->blocksize_1/2;
        float *ret=vc->ret;
        const float *lwin=vc->lwin;
        const float *swin=vc->swin;
        float *buf=vc->buf;
        float *buf_tmp=vc->buf_tmp;

        ch_floor_ptr=vc->channel_floors+j*blocksize/2;

        saved_start=vc->saved_start;

        vc->mdct0.fft.imdct_calc(vc->modes[mode_number].blockflag ? &vc->mdct1 : &vc->mdct0, buf, ch_floor_ptr, buf_tmp);

        //FIXME process channels together, to allow faster simd vector_fmul_add_add?
        if (vc->modes[mode_number].blockflag) {
            // -- overlap/add
            if (previous_window) {
                vc->dsp.vector_fmul_add_add(ret+j, buf, lwin, saved, vc->add_bias, vc->blocksize_1/2, step);
                retlen=vc->blocksize_1/2;
            } else {
                int len = (vc->blocksize_1-vc->blocksize_0)/4;
                buf += len;
                vc->dsp.vector_fmul_add_add(ret+j, buf, swin, saved, vc->add_bias, vc->blocksize_0/2, step);
                k = vc->blocksize_0/2*step + j;
                buf += vc->blocksize_0/2;
                if(vc->exp_bias){
                    for(i=0; i<len; i++, k+=step)
                        ((uint32_t*)ret)[k] = ((uint32_t*)buf)[i] + vc->exp_bias; // ret[k]=buf[i]*(1<<bias)
                } else {
                    for(i=0; i<len; i++, k+=step)
                        ret[k] = buf[i] + fadd_bias;
                }
                buf=vc->buf;
                retlen=vc->blocksize_0/2+len;
            }
            // -- save
            if (next_window) {
                buf += vc->blocksize_1/2;
                vc->dsp.vector_fmul_reverse(saved, buf, lwin, vc->blocksize_1/2);
                saved_start=0;
            } else {
                saved_start=(vc->blocksize_1-vc->blocksize_0)/4;
                buf += vc->blocksize_1/2;
                for(i=0; i<saved_start; i++)
                    ((uint32_t*)saved)[i] = ((uint32_t*)buf)[i] + vc->exp_bias;
                vc->dsp.vector_fmul_reverse(saved+saved_start, buf+saved_start, swin, vc->blocksize_0/2);
            }
        } else {
            // --overlap/add
            if(vc->add_bias) {
                for(k=j, i=0;i<saved_start;++i, k+=step)
                    ret[k] = saved[i] + fadd_bias;
            } else {
                for(k=j, i=0;i<saved_start;++i, k+=step)
                    ret[k] = saved[i];
            }
            vc->dsp.vector_fmul_add_add(ret+k, buf, swin, saved+saved_start, vc->add_bias, vc->blocksize_0/2, step);
            retlen=saved_start+vc->blocksize_0/2;
            // -- save
            buf += vc->blocksize_0/2;
            vc->dsp.vector_fmul_reverse(saved, buf, swin, vc->blocksize_0/2);
            saved_start=0;
        }
    }
    vc->saved_start=saved_start;

    return retlen*vc->audio_channels;
}

// Return the decoded audio packet through the standard api

static int vorbis_decode_frame(AVCodecContext *avccontext,
                        void *data, int *data_size,
                        uint8_t *buf, int buf_size)
{
    vorbis_context *vc = avccontext->priv_data ;
    GetBitContext *gb = &(vc->gb);

    int_fast16_t i, len;

    if(!buf_size){
        return 0;
    }

    AV_DEBUG("packet length %d \n", buf_size);

    init_get_bits(gb, buf, buf_size*8);

    len=vorbis_parse_audio_packet(vc);

    if (len<=0) {
        *data_size=0;
        return buf_size;
    }

    if (!vc->first_frame) {
        vc->first_frame=1;
        *data_size=0;
        return buf_size ;
    }

    AV_DEBUG("parsed %d bytes %d bits, returned %d samples (*ch*bits) \n", get_bits_count(gb)/8, get_bits_count(gb)%8, len);

    vc->dsp.float_to_int16(data, vc->ret, len);
    *data_size=len*2;

    return buf_size ;
}

// Close decoder

static int vorbis_decode_close(AVCodecContext *avccontext) {
    vorbis_context *vc = avccontext->priv_data;

    vorbis_free(vc);

    return 0 ;
}

AVCodec vorbis_decoder = {
    "vorbis",
    CODEC_TYPE_AUDIO,
    CODEC_ID_VORBIS,
    sizeof(vorbis_context),
    vorbis_decode_init,
    NULL,
    vorbis_decode_close,
    vorbis_decode_frame,
};