comparison huffyuv.c @ 903:22ee74da2cd3 libavcodec

cleanup adding AVVideoFrame moving quality, pict_type, key_frame, qscale_table, ... to AVVideoFrame removing obsolete variables in AVCodecContext skiping of MBs in b frames correctly initalizing AVCodecContext picture buffer cleanup

902:6acc8394960d

#ifndef INT64_MAX
#define INT64_MAX 9223372036854775807LL
#endif

#define VLC_BITS 11
  
typedef enum Predictor{
    LEFT= 0,
    PLANE,
    MEDIAN,
} Predictor;

    int bgr32;                              //use bgr32 instead of bgr24
    int width, height;
    int flags;
    int picture_number;
    int last_slice_end;
    int linesize[3];
    uint8_t __align8 temp[3][2500];
    uint64_t stats[3][256];
    uint8_t len[3][256];
    uint32_t bits[3][256];
    VLC vlc[3];
    uint8_t __align8 *picture[3];
    uint8_t __align8 bitstream_buffer[1024*1024*3]; //FIXME dynamic alloc or some other solution
    DSPContext dsp; 
}HYuvContext;

static inline void bswap_buf(uint32_t *dst, uint32_t *src, int w){

}

static int decode_init(AVCodecContext *avctx)
{
    HYuvContext *s = avctx->priv_data;
    int width, height, y_size, c_size, stride;

    s->avctx= avctx;
    s->flags= avctx->flags;
        
    dsputil_init(&s->dsp, avctx->dsp_mask);
    
    width= s->width= avctx->width;
    height= s->height= avctx->height;
s->bgr32=1;
    assert(width && height);
//if(avctx->extradata)
//  printf("extradata:%X, extradata_size:%d\n", *(uint32_t*)avctx->extradata, avctx->extradata_size);
    if(avctx->extradata_size){

            return -1;
    }
    
    s->interlaced= height > 288;
    
    c_size= 0;
    switch(s->bitstream_bpp){
    case 12:
        avctx->pix_fmt = PIX_FMT_YUV420P;
        stride= (width+15)&~15;
        c_size= height*stride/4;
        break;
    case 16:
        if(s->yuy2){
            avctx->pix_fmt = PIX_FMT_YUV422;
            stride= (width*2+15)&~15;
        }else{
            avctx->pix_fmt = PIX_FMT_YUV422P;
            stride= (width+15)&~15;
            c_size= height*stride/2;
        }
        break;
    case 24:
    case 32:
        if(s->bgr32){
            avctx->pix_fmt = PIX_FMT_BGRA32;
            stride= (width*4+15)&~15;
        }else{
            avctx->pix_fmt = PIX_FMT_BGR24;
            stride= (width*3+15)&~15;
        }
        break;
    default:
        assert(0);
        stride=0; //gcc fix
    }
    
    y_size= height*stride;
    
    if(!(avctx->flags&CODEC_FLAG_DR1)){
        s->linesize[0]= stride;
        s->picture[0]= av_mallocz(y_size);
 
        if(c_size){
            s->picture[1]= av_mallocz(c_size);
            s->picture[2]= av_mallocz(c_size);
            s->linesize[1]= s->linesize[2]= stride/2;
        
            memset(s->picture[1], 128, c_size);
            memset(s->picture[2], 128, c_size);
        }
    }
    
//    printf("pred:%d bpp:%d hbpp:%d il:%d\n", s->predictor, s->bitstream_bpp, avctx->bits_per_sample, s->interlaced);
    
    return 0;

    assert(width && height);
    
    avctx->extradata= av_mallocz(1024*10);
    avctx->stats_out= av_mallocz(1024*10);
    s->version=2;
    
    switch(avctx->pix_fmt){
    case PIX_FMT_YUV420P:
        if(avctx->strict_std_compliance>=0){
            fprintf(stderr, "YV12-huffyuv is experimental, there WILL be no compatbility! (use (v)strict=-1)\n");

        cy= y>>1;
    }else{
        cy= y;
    }
    
    src_ptr[0] = s->picture[0] + s->linesize[0]*y;
    src_ptr[1] = s->picture[1] + s->linesize[1]*cy;
    src_ptr[2] = s->picture[2] + s->linesize[2]*cy;
    emms_c();

    s->avctx->draw_horiz_band(s->avctx, src_ptr, s->linesize[0], y, s->width, h);
    
    s->last_slice_end= y + h;
}

static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8_t *buf, int buf_size){
    HYuvContext *s = avctx->priv_data;
    const int width= s->width;
    const int width2= s->width>>1;
    const int height= s->height;
    int fake_ystride, fake_ustride, fake_vstride;
    int i;

    AVPicture *picture = data;

    *data_size = 0;

    /* no supplementary picture */
    if (buf_size == 0)

    bswap_buf((uint32_t*)s->bitstream_buffer, (uint32_t*)buf, buf_size/4);
    
    init_get_bits(&s->gb, s->bitstream_buffer, buf_size);

    if(avctx->flags&CODEC_FLAG_DR1){
        if(avctx->get_buffer_callback(avctx, s->width, s->height, I_TYPE) < 0){
            fprintf(stderr, "get_buffer() failed\n");
            return -1;
        }

        s->linesize[0]= avctx->dr_stride;
        s->linesize[1]=
        s->linesize[2]= avctx->dr_uvstride;

        for(i=0; i<3;i++)
            s->picture[i]= avctx->dr_buffer[i];
    }
    fake_ystride= s->interlaced ? s->linesize[0]*2  : s->linesize[0];
    fake_ustride= s->interlaced ? s->linesize[1]*2  : s->linesize[1];
    fake_vstride= s->interlaced ? s->linesize[2]*2  : s->linesize[2];
    
    s->last_slice_end= 0;
        
    if(s->bitstream_bpp<24){
        int y, cy;
        int lefty, leftu, leftv;
        int lefttopy, lefttopu, lefttopv;
        
        if(s->yuy2){
            s->picture[0][3]= get_bits(&s->gb, 8);
            s->picture[0][2]= get_bits(&s->gb, 8);
            s->picture[0][1]= get_bits(&s->gb, 8);
            s->picture[0][0]= get_bits(&s->gb, 8);
            
            fprintf(stderr, "YUY2 output isnt implemenetd yet\n");
            return -1;
        }else{
        
            leftv= s->picture[2][0]= get_bits(&s->gb, 8);
            lefty= s->picture[0][1]= get_bits(&s->gb, 8);
            leftu= s->picture[1][0]= get_bits(&s->gb, 8);
                   s->picture[0][0]= get_bits(&s->gb, 8);
        
            switch(s->predictor){
            case LEFT:
            case PLANE:
                decode_422_bitstream(s, width-2);
                lefty= add_left_prediction(s->picture[0] + 2, s->temp[0], width-2, lefty);
                if(!(s->flags&CODEC_FLAG_GRAY)){
                    leftu= add_left_prediction(s->picture[1] + 1, s->temp[1], width2-1, leftu);
                    leftv= add_left_prediction(s->picture[2] + 1, s->temp[2], width2-1, leftv);
                }

                for(cy=y=1; y<s->height; y++,cy++){
                    uint8_t *ydst, *udst, *vdst;
                    
                    if(s->bitstream_bpp==12){
                        decode_gray_bitstream(s, width);
                    
                        ydst= s->picture[0] + s->linesize[0]*y;

                        lefty= add_left_prediction(ydst, s->temp[0], width, lefty);
                        if(s->predictor == PLANE){
                            if(y>s->interlaced)
                                s->dsp.add_bytes(ydst, ydst - fake_ystride, width);

                        if(y>=s->height) break;
                    }
                    
                    draw_slice(s, y);
                    
                    ydst= s->picture[0] + s->linesize[0]*y;
                    udst= s->picture[1] + s->linesize[1]*cy;
                    vdst= s->picture[2] + s->linesize[2]*cy;
                    
                    decode_422_bitstream(s, width);

                    lefty= add_left_prediction(ydst, s->temp[0], width, lefty);
                    if(!(s->flags&CODEC_FLAG_GRAY)){
                        leftu= add_left_prediction(udst, s->temp[1], width2, leftu);
                        leftv= add_left_prediction(vdst, s->temp[2], width2, leftv);
                    }

                
                break;
            case MEDIAN:
                /* first line except first 2 pixels is left predicted */
                decode_422_bitstream(s, width-2);
                lefty= add_left_prediction(s->picture[0] + 2, s->temp[0], width-2, lefty);
                if(!(s->flags&CODEC_FLAG_GRAY)){
                    leftu= add_left_prediction(s->picture[1] + 1, s->temp[1], width2-1, leftu);
                    leftv= add_left_prediction(s->picture[2] + 1, s->temp[2], width2-1, leftv);
                }
                
                cy=y=1;
                
                /* second line is left predicted for interlaced case */
                if(s->interlaced){
                    decode_422_bitstream(s, width);
                    lefty= add_left_prediction(s->picture[0] + s->linesize[0], s->temp[0], width, lefty);
                    if(!(s->flags&CODEC_FLAG_GRAY)){
                        leftu= add_left_prediction(s->picture[1] + s->linesize[2], s->temp[1], width2, leftu);
                        leftv= add_left_prediction(s->picture[2] + s->linesize[1], s->temp[2], width2, leftv);
                    }
                    y++; cy++;
                }

                /* next 4 pixels are left predicted too */
                decode_422_bitstream(s, 4);
                lefty= add_left_prediction(s->picture[0] + fake_ystride, s->temp[0], 4, lefty);
                if(!(s->flags&CODEC_FLAG_GRAY)){
                    leftu= add_left_prediction(s->picture[1] + fake_ustride, s->temp[1], 2, leftu);
                    leftv= add_left_prediction(s->picture[2] + fake_vstride, s->temp[2], 2, leftv);
                }

                /* next line except the first 4 pixels is median predicted */
                lefttopy= s->picture[0][3];
                decode_422_bitstream(s, width-4);
                add_median_prediction(s->picture[0] + fake_ystride+4, s->picture[0]+4, s->temp[0], width-4, &lefty, &lefttopy);
                if(!(s->flags&CODEC_FLAG_GRAY)){
                    lefttopu= s->picture[1][1];
                    lefttopv= s->picture[2][1];
                    add_median_prediction(s->picture[1] + fake_ustride+2, s->picture[1]+2, s->temp[1], width2-2, &leftu, &lefttopu);
                    add_median_prediction(s->picture[2] + fake_vstride+2, s->picture[2]+2, s->temp[2], width2-2, &leftv, &lefttopv);
                }
                y++; cy++;
                
                for(; y<height; y++,cy++){
                    uint8_t *ydst, *udst, *vdst;

                    if(s->bitstream_bpp==12){
                        while(2*cy > y){
                            decode_gray_bitstream(s, width);
                            ydst= s->picture[0] + s->linesize[0]*y;
                            add_median_prediction(ydst, ydst - fake_ystride, s->temp[0], width, &lefty, &lefttopy);
                            y++;
                        }
                        if(y>=height) break;
                    }
                    draw_slice(s, y);

                    decode_422_bitstream(s, width);

                    ydst= s->picture[0] + s->linesize[0]*y;
                    udst= s->picture[1] + s->linesize[1]*cy;
                    vdst= s->picture[2] + s->linesize[2]*cy;

                    add_median_prediction(ydst, ydst - fake_ystride, s->temp[0], width, &lefty, &lefttopy);
                    if(!(s->flags&CODEC_FLAG_GRAY)){
                        add_median_prediction(udst, udst - fake_ustride, s->temp[1], width2, &leftu, &lefttopu);
                        add_median_prediction(vdst, vdst - fake_vstride, s->temp[2], width2, &leftv, &lefttopv);

            }
        }
    }else{
        int y;
        int leftr, leftg, leftb;
        const int last_line= (height-1)*s->linesize[0];
        
        if(s->bitstream_bpp==32){
                   s->picture[0][last_line+3]= get_bits(&s->gb, 8);
            leftr= s->picture[0][last_line+2]= get_bits(&s->gb, 8);
            leftg= s->picture[0][last_line+1]= get_bits(&s->gb, 8);
            leftb= s->picture[0][last_line+0]= get_bits(&s->gb, 8);
        }else{
            leftr= s->picture[0][last_line+2]= get_bits(&s->gb, 8);
            leftg= s->picture[0][last_line+1]= get_bits(&s->gb, 8);
            leftb= s->picture[0][last_line+0]= get_bits(&s->gb, 8);
            skip_bits(&s->gb, 8);
        }
        
        if(s->bgr32){
            switch(s->predictor){
            case LEFT:
            case PLANE:
                decode_bgr_bitstream(s, width-1);
                add_left_prediction_bgr32(s->picture[0] + last_line+4, s->temp[0], width-1, &leftr, &leftg, &leftb);

                for(y=s->height-2; y>=0; y--){ //yes its stored upside down
                    decode_bgr_bitstream(s, width);
                    
                    add_left_prediction_bgr32(s->picture[0] + s->linesize[0]*y, s->temp[0], width, &leftr, &leftg, &leftb);
                    if(s->predictor == PLANE){
                        if((y&s->interlaced)==0){
                            s->dsp.add_bytes(s->picture[0] + s->linesize[0]*y, 
                                             s->picture[0] + s->linesize[0]*y + fake_ystride, fake_ystride);
                        }
                    }
                }
                draw_slice(s, height); // just 1 large slice as this isnt possible in reverse order
                break;

            return -1;
        }
    }
    emms_c();
    
    for(i=0;i<3;i++) {
        picture->data[i] = s->picture[i];
        picture->linesize[i]= s->linesize[i];
    }

    *data_size = sizeof(AVPicture);
    
    return (get_bits_count(&s->gb)+7)>>3;
}

static int decode_end(AVCodecContext *avctx)
{
    HYuvContext *s = avctx->priv_data;
    int i;
    
    for(i=0; i<3; i++){
        if(!(avctx->flags&CODEC_FLAG_DR1))
            av_freep(&s->picture[i]);

        free_vlc(&s->vlc[i]);
    }

    return 0;
}

static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data){
    HYuvContext *s = avctx->priv_data;
    AVPicture *pict = data;
    const int width= s->width;
    const int width2= s->width>>1;
    const int height= s->height;
    const int fake_ystride= s->interlaced ? pict->linesize[0]*2  : pict->linesize[0];
    const int fake_ustride= s->interlaced ? pict->linesize[1]*2  : pict->linesize[1];
    const int fake_vstride= s->interlaced ? pict->linesize[2]*2  : pict->linesize[2];
    int i, size;

    init_put_bits(&s->pb, buf, buf_size, NULL, NULL);
    
    for(i=0; i<3; i++){
        s->picture[i]= pict->data[i];
        s->linesize[i]= pict->linesize[i];
    }
    
    if(avctx->pix_fmt == PIX_FMT_YUV422P || avctx->pix_fmt == PIX_FMT_YUV420P){
        int lefty, leftu, leftv, y, cy;

        put_bits(&s->pb, 8, leftv= s->picture[2][0]);
        put_bits(&s->pb, 8, lefty= s->picture[0][1]);
        put_bits(&s->pb, 8, leftu= s->picture[1][0]);
        put_bits(&s->pb, 8,        s->picture[0][0]);
        
        lefty= sub_left_prediction(s, s->temp[0], s->picture[0]+2, width-2 , lefty);
        leftu= sub_left_prediction(s, s->temp[1], s->picture[1]+1, width2-1, leftu);
        leftv= sub_left_prediction(s, s->temp[2], s->picture[2]+1, width2-1, leftv);
        
        encode_422_bitstream(s, width-2);
        
        if(s->predictor==MEDIAN){
            int lefttopy, lefttopu, lefttopv;
            cy=y=1;
            if(s->interlaced){
                lefty= sub_left_prediction(s, s->temp[0], s->picture[0]+s->linesize[0], width , lefty);
                leftu= sub_left_prediction(s, s->temp[1], s->picture[1]+s->linesize[1], width2, leftu);
                leftv= sub_left_prediction(s, s->temp[2], s->picture[2]+s->linesize[2], width2, leftv);
        
                encode_422_bitstream(s, width);
                y++; cy++;
            }
            
            lefty= sub_left_prediction(s, s->temp[0], s->picture[0]+fake_ystride, 4, lefty);
            leftu= sub_left_prediction(s, s->temp[1], s->picture[1]+fake_ystride, 2, leftu);
            leftv= sub_left_prediction(s, s->temp[2], s->picture[2]+fake_ystride, 2, leftv);
        
            encode_422_bitstream(s, 4);

            lefttopy= s->picture[0][3];
            lefttopu= s->picture[1][1];
            lefttopv= s->picture[2][1];
            sub_median_prediction(s->temp[0], s->picture[0]+4, s->picture[0] + fake_ystride+4, width-4 , &lefty, &lefttopy);
            sub_median_prediction(s->temp[1], s->picture[1]+2, s->picture[1] + fake_ustride+2, width2-2, &leftu, &lefttopu);
            sub_median_prediction(s->temp[2], s->picture[2]+2, s->picture[2] + fake_vstride+2, width2-2, &leftv, &lefttopv);
            encode_422_bitstream(s, width-4);
            y++; cy++;

            for(; y<height; y++,cy++){
                uint8_t *ydst, *udst, *vdst;
                    
                if(s->bitstream_bpp==12){
                    while(2*cy > y){
                        ydst= s->picture[0] + s->linesize[0]*y;
                        sub_median_prediction(s->temp[0], ydst - fake_ystride, ydst, width , &lefty, &lefttopy);
                        encode_gray_bitstream(s, width);
                        y++;
                    }
                    if(y>=height) break;
                }
                ydst= s->picture[0] + s->linesize[0]*y;
                udst= s->picture[1] + s->linesize[1]*cy;
                vdst= s->picture[2] + s->linesize[2]*cy;

                sub_median_prediction(s->temp[0], ydst - fake_ystride, ydst, width , &lefty, &lefttopy);
                sub_median_prediction(s->temp[1], udst - fake_ustride, udst, width2, &leftu, &lefttopu);
                sub_median_prediction(s->temp[2], vdst - fake_vstride, vdst, width2, &leftv, &lefttopv);

            for(cy=y=1; y<height; y++,cy++){
                uint8_t *ydst, *udst, *vdst;
                
                /* encode a luma only line & y++ */
                if(s->bitstream_bpp==12){
                    ydst= s->picture[0] + s->linesize[0]*y;

                    if(s->predictor == PLANE && s->interlaced < y){
                        s->dsp.diff_bytes(s->temp[1], ydst, ydst - fake_ystride, width);

                        lefty= sub_left_prediction(s, s->temp[0], s->temp[1], width , lefty);

                    encode_gray_bitstream(s, width);
                    y++;
                    if(y>=height) break;
                }
                
                ydst= s->picture[0] + s->linesize[0]*y;
                udst= s->picture[1] + s->linesize[1]*cy;
                vdst= s->picture[2] + s->linesize[2]*cy;

                if(s->predictor == PLANE && s->interlaced < cy){
                    s->dsp.diff_bytes(s->temp[1], ydst, ydst - fake_ystride, width);
                    s->dsp.diff_bytes(s->temp[2], udst, udst - fake_ustride, width2);
                    s->dsp.diff_bytes(s->temp[3], vdst, vdst - fake_vstride, width2);

    }else{
        flush_put_bits(&s->pb);
        bswap_buf((uint32_t*)buf, (uint32_t*)buf, size);
    }
    
    avctx->key_frame= 1;
    avctx->pict_type= I_TYPE;
    
    s->picture_number++;
    
    return size*4;
}

static int encode_end(AVCodecContext *avctx)
{

903:22ee74da2cd3

#ifndef INT64_MAX
#define INT64_MAX 9223372036854775807LL
#endif

#define VLC_BITS 11

typedef enum Predictor{
    LEFT= 0,
    PLANE,
    MEDIAN,
} Predictor;

    int bgr32;                              //use bgr32 instead of bgr24
    int width, height;
    int flags;
    int picture_number;
    int last_slice_end;
    uint8_t __align8 temp[3][2500];
    uint64_t stats[3][256];
    uint8_t len[3][256];
    uint32_t bits[3][256];
    VLC vlc[3];
    AVVideoFrame picture;
    uint8_t __align8 bitstream_buffer[1024*1024*3]; //FIXME dynamic alloc or some other solution
    DSPContext dsp; 
}HYuvContext;

static inline void bswap_buf(uint32_t *dst, uint32_t *src, int w){

}

static int decode_init(AVCodecContext *avctx)
{
    HYuvContext *s = avctx->priv_data;
    int width, height;

    s->avctx= avctx;
    s->flags= avctx->flags;
        
    dsputil_init(&s->dsp, avctx->dsp_mask);
    
    width= s->width= avctx->width;
    height= s->height= avctx->height;
    avctx->coded_picture= &s->picture;

s->bgr32=1;
    assert(width && height);
//if(avctx->extradata)
//  printf("extradata:%X, extradata_size:%d\n", *(uint32_t*)avctx->extradata, avctx->extradata_size);
    if(avctx->extradata_size){

            return -1;
    }
    
    s->interlaced= height > 288;
    
    switch(s->bitstream_bpp){
    case 12:
        avctx->pix_fmt = PIX_FMT_YUV420P;
        break;
    case 16:
        if(s->yuy2){
            avctx->pix_fmt = PIX_FMT_YUV422;
        }else{
            avctx->pix_fmt = PIX_FMT_YUV422P;
        }
        break;
    case 24:
    case 32:
        if(s->bgr32){
            avctx->pix_fmt = PIX_FMT_BGRA32;
        }else{
            avctx->pix_fmt = PIX_FMT_BGR24;
        }
        break;
    default:
        assert(0);
    }
    
//    printf("pred:%d bpp:%d hbpp:%d il:%d\n", s->predictor, s->bitstream_bpp, avctx->bits_per_sample, s->interlaced);
    
    return 0;

    assert(width && height);
    
    avctx->extradata= av_mallocz(1024*10);
    avctx->stats_out= av_mallocz(1024*10);
    s->version=2;
    
    avctx->coded_picture= &s->picture;
    s->picture.pict_type= I_TYPE;
    s->picture.key_frame= 1;
    
    switch(avctx->pix_fmt){
    case PIX_FMT_YUV420P:
        if(avctx->strict_std_compliance>=0){
            fprintf(stderr, "YV12-huffyuv is experimental, there WILL be no compatbility! (use (v)strict=-1)\n");

        cy= y>>1;
    }else{
        cy= y;
    }
    
    src_ptr[0] = s->picture.data[0] + s->picture.linesize[0]*y;
    src_ptr[1] = s->picture.data[1] + s->picture.linesize[1]*cy;
    src_ptr[2] = s->picture.data[2] + s->picture.linesize[2]*cy;
    emms_c();

    s->avctx->draw_horiz_band(s->avctx, src_ptr, s->picture.linesize[0], y, s->width, h);
    
    s->last_slice_end= y + h;
}

static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8_t *buf, int buf_size){
    HYuvContext *s = avctx->priv_data;
    const int width= s->width;
    const int width2= s->width>>1;
    const int height= s->height;
    int fake_ystride, fake_ustride, fake_vstride;
    AVVideoFrame * const p= &s->picture;

    AVVideoFrame *picture = data;

    *data_size = 0;

    /* no supplementary picture */
    if (buf_size == 0)

    bswap_buf((uint32_t*)s->bitstream_buffer, (uint32_t*)buf, buf_size/4);
    
    init_get_bits(&s->gb, s->bitstream_buffer, buf_size);

    p->reference= 0;
    if(avctx->get_buffer(avctx, p) < 0){
        fprintf(stderr, "get_buffer() failed\n");
        return -1;
    }

    fake_ystride= s->interlaced ? p->linesize[0]*2  : p->linesize[0];
    fake_ustride= s->interlaced ? p->linesize[1]*2  : p->linesize[1];
    fake_vstride= s->interlaced ? p->linesize[2]*2  : p->linesize[2];
    
    s->last_slice_end= 0;
        
    if(s->bitstream_bpp<24){
        int y, cy;
        int lefty, leftu, leftv;
        int lefttopy, lefttopu, lefttopv;
        
        if(s->yuy2){
            p->data[0][3]= get_bits(&s->gb, 8);
            p->data[0][2]= get_bits(&s->gb, 8);
            p->data[0][1]= get_bits(&s->gb, 8);
            p->data[0][0]= get_bits(&s->gb, 8);
            
            fprintf(stderr, "YUY2 output isnt implemenetd yet\n");
            return -1;
        }else{
        
            leftv= p->data[2][0]= get_bits(&s->gb, 8);
            lefty= p->data[0][1]= get_bits(&s->gb, 8);
            leftu= p->data[1][0]= get_bits(&s->gb, 8);
                   p->data[0][0]= get_bits(&s->gb, 8);
        
            switch(s->predictor){
            case LEFT:
            case PLANE:
                decode_422_bitstream(s, width-2);
                lefty= add_left_prediction(p->data[0] + 2, s->temp[0], width-2, lefty);
                if(!(s->flags&CODEC_FLAG_GRAY)){
                    leftu= add_left_prediction(p->data[1] + 1, s->temp[1], width2-1, leftu);
                    leftv= add_left_prediction(p->data[2] + 1, s->temp[2], width2-1, leftv);
                }

                for(cy=y=1; y<s->height; y++,cy++){
                    uint8_t *ydst, *udst, *vdst;
                    
                    if(s->bitstream_bpp==12){
                        decode_gray_bitstream(s, width);
                    
                        ydst= p->data[0] + p->linesize[0]*y;

                        lefty= add_left_prediction(ydst, s->temp[0], width, lefty);
                        if(s->predictor == PLANE){
                            if(y>s->interlaced)
                                s->dsp.add_bytes(ydst, ydst - fake_ystride, width);

                        if(y>=s->height) break;
                    }
                    
                    draw_slice(s, y);
                    
                    ydst= p->data[0] + p->linesize[0]*y;
                    udst= p->data[1] + p->linesize[1]*cy;
                    vdst= p->data[2] + p->linesize[2]*cy;
                    
                    decode_422_bitstream(s, width);
                    lefty= add_left_prediction(ydst, s->temp[0], width, lefty);
                    if(!(s->flags&CODEC_FLAG_GRAY)){
                        leftu= add_left_prediction(udst, s->temp[1], width2, leftu);
                        leftv= add_left_prediction(vdst, s->temp[2], width2, leftv);
                    }

                
                break;
            case MEDIAN:
                /* first line except first 2 pixels is left predicted */
                decode_422_bitstream(s, width-2);
                lefty= add_left_prediction(p->data[0] + 2, s->temp[0], width-2, lefty);
                if(!(s->flags&CODEC_FLAG_GRAY)){
                    leftu= add_left_prediction(p->data[1] + 1, s->temp[1], width2-1, leftu);
                    leftv= add_left_prediction(p->data[2] + 1, s->temp[2], width2-1, leftv);
                }
                
                cy=y=1;
                
                /* second line is left predicted for interlaced case */
                if(s->interlaced){
                    decode_422_bitstream(s, width);
                    lefty= add_left_prediction(p->data[0] + p->linesize[0], s->temp[0], width, lefty);
                    if(!(s->flags&CODEC_FLAG_GRAY)){
                        leftu= add_left_prediction(p->data[1] + p->linesize[2], s->temp[1], width2, leftu);
                        leftv= add_left_prediction(p->data[2] + p->linesize[1], s->temp[2], width2, leftv);
                    }
                    y++; cy++;
                }

                /* next 4 pixels are left predicted too */
                decode_422_bitstream(s, 4);
                lefty= add_left_prediction(p->data[0] + fake_ystride, s->temp[0], 4, lefty);
                if(!(s->flags&CODEC_FLAG_GRAY)){
                    leftu= add_left_prediction(p->data[1] + fake_ustride, s->temp[1], 2, leftu);
                    leftv= add_left_prediction(p->data[2] + fake_vstride, s->temp[2], 2, leftv);
                }

                /* next line except the first 4 pixels is median predicted */
                lefttopy= p->data[0][3];
                decode_422_bitstream(s, width-4);
                add_median_prediction(p->data[0] + fake_ystride+4, p->data[0]+4, s->temp[0], width-4, &lefty, &lefttopy);
                if(!(s->flags&CODEC_FLAG_GRAY)){
                    lefttopu= p->data[1][1];
                    lefttopv= p->data[2][1];
                    add_median_prediction(p->data[1] + fake_ustride+2, p->data[1]+2, s->temp[1], width2-2, &leftu, &lefttopu);
                    add_median_prediction(p->data[2] + fake_vstride+2, p->data[2]+2, s->temp[2], width2-2, &leftv, &lefttopv);
                }
                y++; cy++;
                
                for(; y<height; y++,cy++){
                    uint8_t *ydst, *udst, *vdst;

                    if(s->bitstream_bpp==12){
                        while(2*cy > y){
                            decode_gray_bitstream(s, width);
                            ydst= p->data[0] + p->linesize[0]*y;
                            add_median_prediction(ydst, ydst - fake_ystride, s->temp[0], width, &lefty, &lefttopy);
                            y++;
                        }
                        if(y>=height) break;
                    }
                    draw_slice(s, y);

                    decode_422_bitstream(s, width);

                    ydst= p->data[0] + p->linesize[0]*y;
                    udst= p->data[1] + p->linesize[1]*cy;
                    vdst= p->data[2] + p->linesize[2]*cy;

                    add_median_prediction(ydst, ydst - fake_ystride, s->temp[0], width, &lefty, &lefttopy);
                    if(!(s->flags&CODEC_FLAG_GRAY)){
                        add_median_prediction(udst, udst - fake_ustride, s->temp[1], width2, &leftu, &lefttopu);
                        add_median_prediction(vdst, vdst - fake_vstride, s->temp[2], width2, &leftv, &lefttopv);

            }
        }
    }else{
        int y;
        int leftr, leftg, leftb;
        const int last_line= (height-1)*p->linesize[0];
        
        if(s->bitstream_bpp==32){
                   p->data[0][last_line+3]= get_bits(&s->gb, 8);
            leftr= p->data[0][last_line+2]= get_bits(&s->gb, 8);
            leftg= p->data[0][last_line+1]= get_bits(&s->gb, 8);
            leftb= p->data[0][last_line+0]= get_bits(&s->gb, 8);
        }else{
            leftr= p->data[0][last_line+2]= get_bits(&s->gb, 8);
            leftg= p->data[0][last_line+1]= get_bits(&s->gb, 8);
            leftb= p->data[0][last_line+0]= get_bits(&s->gb, 8);
            skip_bits(&s->gb, 8);
        }
        
        if(s->bgr32){
            switch(s->predictor){
            case LEFT:
            case PLANE:
                decode_bgr_bitstream(s, width-1);
                add_left_prediction_bgr32(p->data[0] + last_line+4, s->temp[0], width-1, &leftr, &leftg, &leftb);

                for(y=s->height-2; y>=0; y--){ //yes its stored upside down
                    decode_bgr_bitstream(s, width);
                    
                    add_left_prediction_bgr32(p->data[0] + p->linesize[0]*y, s->temp[0], width, &leftr, &leftg, &leftb);
                    if(s->predictor == PLANE){
                        if((y&s->interlaced)==0){
                            s->dsp.add_bytes(p->data[0] + p->linesize[0]*y, 
                                             p->data[0] + p->linesize[0]*y + fake_ystride, fake_ystride);
                        }
                    }
                }
                draw_slice(s, height); // just 1 large slice as this isnt possible in reverse order
                break;

            return -1;
        }
    }
    emms_c();
    
    *picture= *p;
    
    avctx->release_buffer(avctx, p);

    *data_size = sizeof(AVVideoFrame);
    
    return (get_bits_count(&s->gb)+7)>>3;
}

static int decode_end(AVCodecContext *avctx)
{
    HYuvContext *s = avctx->priv_data;
    int i;
    
    for(i=0; i<3; i++){
        free_vlc(&s->vlc[i]);
    }

    if(avctx->get_buffer == avcodec_default_get_buffer){
        for(i=0; i<4; i++){
            av_freep(&s->picture.base[i]);
            s->picture.data[i]= NULL;
        }
        av_freep(&s->picture.opaque);
    }

    return 0;
}

static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data){
    HYuvContext *s = avctx->priv_data;
    AVVideoFrame *pict = data;
    const int width= s->width;
    const int width2= s->width>>1;
    const int height= s->height;
    const int fake_ystride= s->interlaced ? pict->linesize[0]*2  : pict->linesize[0];
    const int fake_ustride= s->interlaced ? pict->linesize[1]*2  : pict->linesize[1];
    const int fake_vstride= s->interlaced ? pict->linesize[2]*2  : pict->linesize[2];
    AVVideoFrame * const p= &s->picture;
    int i, size;

    init_put_bits(&s->pb, buf, buf_size, NULL, NULL);
    
    *p = *pict;
    
    if(avctx->pix_fmt == PIX_FMT_YUV422P || avctx->pix_fmt == PIX_FMT_YUV420P){
        int lefty, leftu, leftv, y, cy;

        put_bits(&s->pb, 8, leftv= p->data[2][0]);
        put_bits(&s->pb, 8, lefty= p->data[0][1]);
        put_bits(&s->pb, 8, leftu= p->data[1][0]);
        put_bits(&s->pb, 8,        p->data[0][0]);
        
        lefty= sub_left_prediction(s, s->temp[0], p->data[0]+2, width-2 , lefty);
        leftu= sub_left_prediction(s, s->temp[1], p->data[1]+1, width2-1, leftu);
        leftv= sub_left_prediction(s, s->temp[2], p->data[2]+1, width2-1, leftv);
        
        encode_422_bitstream(s, width-2);
        
        if(s->predictor==MEDIAN){
            int lefttopy, lefttopu, lefttopv;
            cy=y=1;
            if(s->interlaced){
                lefty= sub_left_prediction(s, s->temp[0], p->data[0]+p->linesize[0], width , lefty);
                leftu= sub_left_prediction(s, s->temp[1], p->data[1]+p->linesize[1], width2, leftu);
                leftv= sub_left_prediction(s, s->temp[2], p->data[2]+p->linesize[2], width2, leftv);
        
                encode_422_bitstream(s, width);
                y++; cy++;
            }
            
            lefty= sub_left_prediction(s, s->temp[0], p->data[0]+fake_ystride, 4, lefty);
            leftu= sub_left_prediction(s, s->temp[1], p->data[1]+fake_ystride, 2, leftu);
            leftv= sub_left_prediction(s, s->temp[2], p->data[2]+fake_ystride, 2, leftv);
        
            encode_422_bitstream(s, 4);

            lefttopy= p->data[0][3];
            lefttopu= p->data[1][1];
            lefttopv= p->data[2][1];
            sub_median_prediction(s->temp[0], p->data[0]+4, p->data[0] + fake_ystride+4, width-4 , &lefty, &lefttopy);
            sub_median_prediction(s->temp[1], p->data[1]+2, p->data[1] + fake_ustride+2, width2-2, &leftu, &lefttopu);
            sub_median_prediction(s->temp[2], p->data[2]+2, p->data[2] + fake_vstride+2, width2-2, &leftv, &lefttopv);
            encode_422_bitstream(s, width-4);
            y++; cy++;

            for(; y<height; y++,cy++){
                uint8_t *ydst, *udst, *vdst;
                    
                if(s->bitstream_bpp==12){
                    while(2*cy > y){
                        ydst= p->data[0] + p->linesize[0]*y;
                        sub_median_prediction(s->temp[0], ydst - fake_ystride, ydst, width , &lefty, &lefttopy);
                        encode_gray_bitstream(s, width);
                        y++;
                    }
                    if(y>=height) break;
                }
                ydst= p->data[0] + p->linesize[0]*y;
                udst= p->data[1] + p->linesize[1]*cy;
                vdst= p->data[2] + p->linesize[2]*cy;

                sub_median_prediction(s->temp[0], ydst - fake_ystride, ydst, width , &lefty, &lefttopy);
                sub_median_prediction(s->temp[1], udst - fake_ustride, udst, width2, &leftu, &lefttopu);
                sub_median_prediction(s->temp[2], vdst - fake_vstride, vdst, width2, &leftv, &lefttopv);

            for(cy=y=1; y<height; y++,cy++){
                uint8_t *ydst, *udst, *vdst;
                
                /* encode a luma only line & y++ */
                if(s->bitstream_bpp==12){
                    ydst= p->data[0] + p->linesize[0]*y;

                    if(s->predictor == PLANE && s->interlaced < y){
                        s->dsp.diff_bytes(s->temp[1], ydst, ydst - fake_ystride, width);

                        lefty= sub_left_prediction(s, s->temp[0], s->temp[1], width , lefty);

                    encode_gray_bitstream(s, width);
                    y++;
                    if(y>=height) break;
                }
                
                ydst= p->data[0] + p->linesize[0]*y;
                udst= p->data[1] + p->linesize[1]*cy;
                vdst= p->data[2] + p->linesize[2]*cy;

                if(s->predictor == PLANE && s->interlaced < cy){
                    s->dsp.diff_bytes(s->temp[1], ydst, ydst - fake_ystride, width);
                    s->dsp.diff_bytes(s->temp[2], udst, udst - fake_ustride, width2);
                    s->dsp.diff_bytes(s->temp[3], vdst, vdst - fake_vstride, width2);

    }else{
        flush_put_bits(&s->pb);
        bswap_buf((uint32_t*)buf, (uint32_t*)buf, size);
    }
    
    s->picture_number++;

    return size*4;
}

static int encode_end(AVCodecContext *avctx)
{