comparison snow.c @ 6405:7f615bfc21a2 libavcodec

Remove START_TIMER and STOP_TIMER macros.

6404:a0a1db4738dd

    for(y=-2; y<height; y+=2){
        DWTELEM *b2= buffer + mirror(y+1, height-1)*stride;
        DWTELEM *b3= buffer + mirror(y+2, height-1)*stride;

{START_TIMER
        if(y+1<(unsigned)height) horizontal_decompose53i(b2, width);
        if(y+2<(unsigned)height) horizontal_decompose53i(b3, width);
STOP_TIMER("horizontal_decompose53i")}

{START_TIMER
        if(y+1<(unsigned)height) vertical_decompose53iH0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_decompose53iL0(b0, b1, b2, width);
STOP_TIMER("vertical_decompose53i*")}

        b0=b2;
        b1=b3;
    }
}

    for(y=-4; y<height; y+=2){
        DWTELEM *b4= buffer + mirror(y+3, height-1)*stride;
        DWTELEM *b5= buffer + mirror(y+4, height-1)*stride;

{START_TIMER
        if(y+3<(unsigned)height) horizontal_decompose97i(b4, width);
        if(y+4<(unsigned)height) horizontal_decompose97i(b5, width);
if(width>400){
STOP_TIMER("horizontal_decompose97i")
}}

{START_TIMER
        if(y+3<(unsigned)height) vertical_decompose97iH0(b3, b4, b5, width);
        if(y+2<(unsigned)height) vertical_decompose97iL0(b2, b3, b4, width);
        if(y+1<(unsigned)height) vertical_decompose97iH1(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_decompose97iL1(b0, b1, b2, width);

if(width>400){
STOP_TIMER("vertical_decompose97i")
}}

        b0=b2;
        b1=b3;
        b2=b4;
        b3=b5;

    IDWTELEM *b0= cs->b0;
    IDWTELEM *b1= cs->b1;
    IDWTELEM *b2= slice_buffer_get_line(sb, mirror(y+1, height-1) * stride_line);
    IDWTELEM *b3= slice_buffer_get_line(sb, mirror(y+2, height-1) * stride_line);

{START_TIMER
        if(y+1<(unsigned)height) vertical_compose53iL0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_compose53iH0(b0, b1, b2, width);
STOP_TIMER("vertical_compose53i*")}

{START_TIMER
        if(y-1<(unsigned)height) horizontal_compose53i(b0, width);
        if(y+0<(unsigned)height) horizontal_compose53i(b1, width);
STOP_TIMER("horizontal_compose53i")}

    cs->b0 = b2;
    cs->b1 = b3;
    cs->y += 2;
}

    IDWTELEM *b0= cs->b0;
    IDWTELEM *b1= cs->b1;
    IDWTELEM *b2= buffer + mirror(y+1, height-1)*stride;
    IDWTELEM *b3= buffer + mirror(y+2, height-1)*stride;

{START_TIMER
        if(y+1<(unsigned)height) vertical_compose53iL0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_compose53iH0(b0, b1, b2, width);
STOP_TIMER("vertical_compose53i*")}

{START_TIMER
        if(y-1<(unsigned)height) horizontal_compose53i(b0, width);
        if(y+0<(unsigned)height) horizontal_compose53i(b1, width);
STOP_TIMER("horizontal_compose53i")}

    cs->b0 = b2;
    cs->b1 = b3;
    cs->y += 2;
}

    IDWTELEM *b2= cs->b2;
    IDWTELEM *b3= cs->b3;
    IDWTELEM *b4= slice_buffer_get_line(sb, mirror(y + 3, height - 1) * stride_line);
    IDWTELEM *b5= slice_buffer_get_line(sb, mirror(y + 4, height - 1) * stride_line);

{START_TIMER
    if(y>0 && y+4<height){
        dsp->vertical_compose97i(b0, b1, b2, b3, b4, b5, width);
    }else{
        if(y+3<(unsigned)height) vertical_compose97iL1(b3, b4, b5, width);
        if(y+2<(unsigned)height) vertical_compose97iH1(b2, b3, b4, width);
        if(y+1<(unsigned)height) vertical_compose97iL0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_compose97iH0(b0, b1, b2, width);
    }
if(width>400){
STOP_TIMER("vertical_compose97i")}}

{START_TIMER
        if(y-1<(unsigned)height) dsp->horizontal_compose97i(b0, width);
        if(y+0<(unsigned)height) dsp->horizontal_compose97i(b1, width);
if(width>400 && y+0<(unsigned)height){
STOP_TIMER("horizontal_compose97i")}}

    cs->b0=b2;
    cs->b1=b3;
    cs->b2=b4;
    cs->b3=b5;

    IDWTELEM *b2= cs->b2;
    IDWTELEM *b3= cs->b3;
    IDWTELEM *b4= buffer + mirror(y+3, height-1)*stride;
    IDWTELEM *b5= buffer + mirror(y+4, height-1)*stride;

{START_TIMER
        if(y+3<(unsigned)height) vertical_compose97iL1(b3, b4, b5, width);
        if(y+2<(unsigned)height) vertical_compose97iH1(b2, b3, b4, width);
        if(y+1<(unsigned)height) vertical_compose97iL0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_compose97iH0(b0, b1, b2, width);
if(width>400){
STOP_TIMER("vertical_compose97i")}}

{START_TIMER
        if(y-1<(unsigned)height) ff_snow_horizontal_compose97i(b0, width);
        if(y+0<(unsigned)height) ff_snow_horizontal_compose97i(b1, width);
if(width>400 && b0 <= b2){
STOP_TIMER("horizontal_compose97i")}}

    cs->b0=b2;
    cs->b1=b3;
    cs->b2=b4;
    cs->b3=b5;

    const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
    int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
    int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
    int new_index = 0;

    START_TIMER

    if(b->ibuf == s->spatial_idwt_buffer || s->qlog == LOSSLESS_QLOG){
        qadd= 0;
        qmul= 1<<QEXPSHIFT;
    }

            line[x] = (t^u) - u;

            v = b->x_coeff[new_index].coeff;
            x = b->x_coeff[new_index++].x;
        }
    }
    if(w > 200 && start_y != 0/*level+1 == s->spatial_decomposition_count*/){
        STOP_TIMER("decode_subband")
    }

    /* Save our variables for the next slice. */
    save_state[0] = new_index;

    int16_t tmpIt   [64*(32+HTAPS_MAX)];
    uint8_t tmp2t[3][stride*(32+HTAPS_MAX)];
    int16_t *tmpI= tmpIt;
    uint8_t *tmp2= tmp2t[0];
    const uint8_t *hpel[11];
START_TIMER
    assert(dx<16 && dy<16);
    r= brane[dx + 16*dy]&15;
    l= brane[dx + 16*dy]>>4;

    b= needs[l] | needs[r];

            src1+=stride;
            src2+=stride;
            dst +=stride;
        }
    }
STOP_TIMER("mc_block")
}

#define mca(dx,dy,b_w)\
static void mc_block_hpel ## dx ## dy ## b_w(uint8_t *dst, const uint8_t *src, int stride, int h){\
    uint8_t tmp[stride*(b_w+HTAPS_MAX-1)];\

            else    dst[x + y*dst_stride] -= v;
        }
    }
#else
    if(sliced){
        START_TIMER

        s->dsp.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
        STOP_TIMER("inner_add_yblock")
    }else
    for(y=0; y<b_h; y++){
        //FIXME ugly misuse of obmc_stride
        const uint8_t *obmc1= obmc + y*obmc_stride;
        const uint8_t *obmc2= obmc1+ (obmc_stride>>1);

    int obmc_stride= plane_index ? block_size : 2*block_size;
    int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *dst8= s->current_picture.data[plane_index];
    int w= p->width;
    int h= p->height;
    START_TIMER

    if(s->keyframe || (s->avctx->debug&512)){
        if(mb_y==mb_h)
            return;

        return;
    }

        for(mb_x=0; mb_x<=mb_w; mb_x++){
            START_TIMER

            add_yblock(s, 1, sb, old_buffer, dst8, obmc,
                       block_w*mb_x - block_w/2,
                       block_w*mb_y - block_w/2,
                       block_w, block_w,
                       w, h,
                       w, ref_stride, obmc_stride,
                       mb_x - 1, mb_y - 1,
                       add, 0, plane_index);

            STOP_TIMER("add_yblock")
        }

    STOP_TIMER("predict_slice")
}

static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
    Plane *p= &s->plane[plane_index];
    const int mb_w= s->b_width  << s->block_max_depth;

    const int obmc_stride= plane_index ? block_size : 2*block_size;
    int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *dst8= s->current_picture.data[plane_index];
    int w= p->width;
    int h= p->height;
    START_TIMER

    if(s->keyframe || (s->avctx->debug&512)){
        if(mb_y==mb_h)
            return;

        return;
    }

        for(mb_x=0; mb_x<=mb_w; mb_x++){
            START_TIMER

            add_yblock(s, 0, NULL, buf, dst8, obmc,
                       block_w*mb_x - block_w/2,
                       block_w*mb_y - block_w/2,
                       block_w, block_w,
                       w, h,
                       w, ref_stride, obmc_stride,
                       mb_x - 1, mb_y - 1,
                       add, 1, plane_index);

            STOP_TIMER("add_yblock")
        }

    STOP_TIMER("predict_slice")
}

static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
    const int mb_h= s->b_height << s->block_max_depth;
    int mb_y;

        av_log(NULL, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
    }
}

static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
    const int level= b->level;
    const int w= b->width;
    const int h= b->height;
    const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
    const int qmul= qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
    int x,y, thres1, thres2;
//    START_TIMER

    if(s->qlog == LOSSLESS_QLOG){
        for(y=0; y<h; y++)
            for(x=0; x<w; x++)
                dst[x + y*stride]= src[x + y*stride];

                }else
                    dst[x + y*stride]= 0;
            }
        }
    }
    if(level+1 == s->spatial_decomposition_count){
//        STOP_TIMER("quantize")
    }
}

static void dequantize_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, IDWTELEM *src, int stride, int start_y, int end_y){
    const int w= b->width;
    const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
    const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
    const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
    int x,y;
    START_TIMER

    if(s->qlog == LOSSLESS_QLOG) return;

    for(y=start_y; y<end_y; y++){
//        DWTELEM * line = slice_buffer_get_line_from_address(sb, src + (y * stride));

            }else if(i>0){
                line[x]=  (( i*qmul + qadd)>>(QEXPSHIFT));
            }
        }
    }
    if(w > 200 /*level+1 == s->spatial_decomposition_count*/){
        STOP_TIMER("dquant")
    }
}

static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
    const int w= b->width;
    const int h= b->height;
    const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
    const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
    const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
    int x,y;
    START_TIMER

    if(s->qlog == LOSSLESS_QLOG) return;

    for(y=0; y<h; y++){
        for(x=0; x<w; x++){

                src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
            }else if(i>0){
                src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
            }
        }
    }
    if(w > 200 /*level+1 == s->spatial_decomposition_count*/){
        STOP_TIMER("dquant")
    }
}

static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
    const int w= b->width;

}

static void correlate_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median, int start_y, int end_y){
    const int w= b->width;
    int x,y;

//    START_TIMER

    IDWTELEM * line=0; // silence silly "could be used without having been initialized" warning
    IDWTELEM * prev;

    if (start_y != 0)

            }else{
                if(y) line[x] += prev[x];
            }
        }
    }

//    STOP_TIMER("correlate")
}

static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
    const int w= b->width;
    const int h= b->height;

                for(x=0; x<w; x++){
                    s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
                }
            }
        }
{START_TIMER
        predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
STOP_TIMER("pred-conv")}
      }else{
            //ME/MC only
            if(pict->pict_type == I_TYPE){
                for(y=0; y<h; y++){
                    for(x=0; x<w; x++){

                s->mconly_picture.data[plane_index][y*s->mconly_picture.linesize[plane_index] + x]= v;
            }
        }
}

{   START_TIMER
    for(level=0; level<s->spatial_decomposition_count; level++){
        for(orientation=level ? 1 : 0; orientation<4; orientation++){
            SubBand *b= &p->band[level][orientation];
            unpack_coeffs(s, b, b->parent, orientation);
        }
    }
    STOP_TIMER("unpack coeffs");
}

{START_TIMER
    const int mb_h= s->b_height << s->block_max_depth;
    const int block_size = MB_SIZE >> s->block_max_depth;
    const int block_w    = plane_index ? block_size/2 : block_size;
    int mb_y;
    dwt_compose_t cs[MAX_DECOMPOSITIONS];

        if (!(s->keyframe || s->avctx->debug&512)){
            slice_starty = FFMAX(0, slice_starty - (block_w >> 1));
            slice_h -= (block_w >> 1);
        }

        {
        START_TIMER
        for(level=0; level<s->spatial_decomposition_count; level++){
            for(orientation=level ? 1 : 0; orientation<4; orientation++){
                SubBand *b= &p->band[level][orientation];
                int start_y;
                int end_y;

                    else
                        decode_subband_slice_buffered(s, b, &s->sb, start_y, end_y, decode_state[level][orientation]);
                }
            }
        }
        STOP_TIMER("decode_subband_slice");
        }

{   START_TIMER
        for(; yd<slice_h; yd+=4){
            ff_spatial_idwt_buffered_slice(&s->dsp, cs, &s->sb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count, yd);
        }
    STOP_TIMER("idwt slice");}


        if(s->qlog == LOSSLESS_QLOG){
            for(; yq<slice_h && yq<h; yq++){
                IDWTELEM * line = slice_buffer_get_line(&s->sb, yq);
                for(x=0; x<w; x++){

        while(y < end_y)
            slice_buffer_release(&s->sb, y++);
    }

    slice_buffer_flush(&s->sb);

STOP_TIMER("idwt + predict_slices")}
    }

    emms_c();

    if(s->last_picture[s->max_ref_frames-1].data[0]){

    memset(s.header_state, 0, sizeof(s.header_state));
    ff_init_range_encoder(&s.c, buffer[0], 256*256);
    ff_init_cabac_states(&s.c, ff_h264_lps_range, ff_h264_mps_state, ff_h264_lps_state, 64);

    for(i=-256; i<256; i++){
START_TIMER
        put_symbol(&s.c, s.header_state, i*i*i/3*FFABS(i), 1);
STOP_TIMER("put_symbol")
    }
    ff_rac_terminate(&s.c);

    memset(s.header_state, 0, sizeof(s.header_state));
    ff_init_range_decoder(&s.c, buffer[0], 256*256);
    ff_init_cabac_states(&s.c, ff_h264_lps_range, ff_h264_mps_state, ff_h264_lps_state, 64);

    for(i=-256; i<256; i++){
        int j;
START_TIMER
        j= get_symbol(&s.c, s.header_state, 1);
STOP_TIMER("get_symbol")
        if(j!=i*i*i/3*FFABS(i)) printf("fsck: %d != %d\n", i, j);
    }
#endif
{
int level, orientation, x, y;

6405:7f615bfc21a2

    for(y=-2; y<height; y+=2){
        DWTELEM *b2= buffer + mirror(y+1, height-1)*stride;
        DWTELEM *b3= buffer + mirror(y+2, height-1)*stride;

        if(y+1<(unsigned)height) horizontal_decompose53i(b2, width);
        if(y+2<(unsigned)height) horizontal_decompose53i(b3, width);

        if(y+1<(unsigned)height) vertical_decompose53iH0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_decompose53iL0(b0, b1, b2, width);

        b0=b2;
        b1=b3;
    }
}

    for(y=-4; y<height; y+=2){
        DWTELEM *b4= buffer + mirror(y+3, height-1)*stride;
        DWTELEM *b5= buffer + mirror(y+4, height-1)*stride;

        if(y+3<(unsigned)height) horizontal_decompose97i(b4, width);
        if(y+4<(unsigned)height) horizontal_decompose97i(b5, width);

        if(y+3<(unsigned)height) vertical_decompose97iH0(b3, b4, b5, width);
        if(y+2<(unsigned)height) vertical_decompose97iL0(b2, b3, b4, width);
        if(y+1<(unsigned)height) vertical_decompose97iH1(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_decompose97iL1(b0, b1, b2, width);

        b0=b2;
        b1=b3;
        b2=b4;
        b3=b5;

    IDWTELEM *b0= cs->b0;
    IDWTELEM *b1= cs->b1;
    IDWTELEM *b2= slice_buffer_get_line(sb, mirror(y+1, height-1) * stride_line);
    IDWTELEM *b3= slice_buffer_get_line(sb, mirror(y+2, height-1) * stride_line);

        if(y+1<(unsigned)height) vertical_compose53iL0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_compose53iH0(b0, b1, b2, width);

        if(y-1<(unsigned)height) horizontal_compose53i(b0, width);
        if(y+0<(unsigned)height) horizontal_compose53i(b1, width);

    cs->b0 = b2;
    cs->b1 = b3;
    cs->y += 2;
}

    IDWTELEM *b0= cs->b0;
    IDWTELEM *b1= cs->b1;
    IDWTELEM *b2= buffer + mirror(y+1, height-1)*stride;
    IDWTELEM *b3= buffer + mirror(y+2, height-1)*stride;

        if(y+1<(unsigned)height) vertical_compose53iL0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_compose53iH0(b0, b1, b2, width);

        if(y-1<(unsigned)height) horizontal_compose53i(b0, width);
        if(y+0<(unsigned)height) horizontal_compose53i(b1, width);

    cs->b0 = b2;
    cs->b1 = b3;
    cs->y += 2;
}

    IDWTELEM *b2= cs->b2;
    IDWTELEM *b3= cs->b3;
    IDWTELEM *b4= slice_buffer_get_line(sb, mirror(y + 3, height - 1) * stride_line);
    IDWTELEM *b5= slice_buffer_get_line(sb, mirror(y + 4, height - 1) * stride_line);

    if(y>0 && y+4<height){
        dsp->vertical_compose97i(b0, b1, b2, b3, b4, b5, width);
    }else{
        if(y+3<(unsigned)height) vertical_compose97iL1(b3, b4, b5, width);
        if(y+2<(unsigned)height) vertical_compose97iH1(b2, b3, b4, width);
        if(y+1<(unsigned)height) vertical_compose97iL0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_compose97iH0(b0, b1, b2, width);
    }

        if(y-1<(unsigned)height) dsp->horizontal_compose97i(b0, width);
        if(y+0<(unsigned)height) dsp->horizontal_compose97i(b1, width);

    cs->b0=b2;
    cs->b1=b3;
    cs->b2=b4;
    cs->b3=b5;

    IDWTELEM *b2= cs->b2;
    IDWTELEM *b3= cs->b3;
    IDWTELEM *b4= buffer + mirror(y+3, height-1)*stride;
    IDWTELEM *b5= buffer + mirror(y+4, height-1)*stride;

        if(y+3<(unsigned)height) vertical_compose97iL1(b3, b4, b5, width);
        if(y+2<(unsigned)height) vertical_compose97iH1(b2, b3, b4, width);
        if(y+1<(unsigned)height) vertical_compose97iL0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_compose97iH0(b0, b1, b2, width);

        if(y-1<(unsigned)height) ff_snow_horizontal_compose97i(b0, width);
        if(y+0<(unsigned)height) ff_snow_horizontal_compose97i(b1, width);

    cs->b0=b2;
    cs->b1=b3;
    cs->b2=b4;
    cs->b3=b5;

    const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
    int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
    int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
    int new_index = 0;

    if(b->ibuf == s->spatial_idwt_buffer || s->qlog == LOSSLESS_QLOG){
        qadd= 0;
        qmul= 1<<QEXPSHIFT;
    }

            line[x] = (t^u) - u;

            v = b->x_coeff[new_index].coeff;
            x = b->x_coeff[new_index++].x;
        }
    }

    /* Save our variables for the next slice. */
    save_state[0] = new_index;

    int16_t tmpIt   [64*(32+HTAPS_MAX)];
    uint8_t tmp2t[3][stride*(32+HTAPS_MAX)];
    int16_t *tmpI= tmpIt;
    uint8_t *tmp2= tmp2t[0];
    const uint8_t *hpel[11];
    assert(dx<16 && dy<16);
    r= brane[dx + 16*dy]&15;
    l= brane[dx + 16*dy]>>4;

    b= needs[l] | needs[r];

            src1+=stride;
            src2+=stride;
            dst +=stride;
        }
    }
}

#define mca(dx,dy,b_w)\
static void mc_block_hpel ## dx ## dy ## b_w(uint8_t *dst, const uint8_t *src, int stride, int h){\
    uint8_t tmp[stride*(b_w+HTAPS_MAX-1)];\

            else    dst[x + y*dst_stride] -= v;
        }
    }
#else
    if(sliced){
        s->dsp.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
    }else
    for(y=0; y<b_h; y++){
        //FIXME ugly misuse of obmc_stride
        const uint8_t *obmc1= obmc + y*obmc_stride;
        const uint8_t *obmc2= obmc1+ (obmc_stride>>1);

    int obmc_stride= plane_index ? block_size : 2*block_size;
    int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *dst8= s->current_picture.data[plane_index];
    int w= p->width;
    int h= p->height;

    if(s->keyframe || (s->avctx->debug&512)){
        if(mb_y==mb_h)
            return;

        return;
    }

        for(mb_x=0; mb_x<=mb_w; mb_x++){
            add_yblock(s, 1, sb, old_buffer, dst8, obmc,
                       block_w*mb_x - block_w/2,
                       block_w*mb_y - block_w/2,
                       block_w, block_w,
                       w, h,
                       w, ref_stride, obmc_stride,
                       mb_x - 1, mb_y - 1,
                       add, 0, plane_index);
        }
}

static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
    Plane *p= &s->plane[plane_index];
    const int mb_w= s->b_width  << s->block_max_depth;

    const int obmc_stride= plane_index ? block_size : 2*block_size;
    int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *dst8= s->current_picture.data[plane_index];
    int w= p->width;
    int h= p->height;

    if(s->keyframe || (s->avctx->debug&512)){
        if(mb_y==mb_h)
            return;

        return;
    }

        for(mb_x=0; mb_x<=mb_w; mb_x++){
            add_yblock(s, 0, NULL, buf, dst8, obmc,
                       block_w*mb_x - block_w/2,
                       block_w*mb_y - block_w/2,
                       block_w, block_w,
                       w, h,
                       w, ref_stride, obmc_stride,
                       mb_x - 1, mb_y - 1,
                       add, 1, plane_index);
        }
}

static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
    const int mb_h= s->b_height << s->block_max_depth;
    int mb_y;

        av_log(NULL, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
    }
}

static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
    const int w= b->width;
    const int h= b->height;
    const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
    const int qmul= qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
    int x,y, thres1, thres2;

    if(s->qlog == LOSSLESS_QLOG){
        for(y=0; y<h; y++)
            for(x=0; x<w; x++)
                dst[x + y*stride]= src[x + y*stride];

                }else
                    dst[x + y*stride]= 0;
            }
        }
    }
}

static void dequantize_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, IDWTELEM *src, int stride, int start_y, int end_y){
    const int w= b->width;
    const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
    const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
    const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
    int x,y;

    if(s->qlog == LOSSLESS_QLOG) return;

    for(y=start_y; y<end_y; y++){
//        DWTELEM * line = slice_buffer_get_line_from_address(sb, src + (y * stride));

            }else if(i>0){
                line[x]=  (( i*qmul + qadd)>>(QEXPSHIFT));
            }
        }
    }
}

static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
    const int w= b->width;
    const int h= b->height;
    const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
    const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
    const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
    int x,y;

    if(s->qlog == LOSSLESS_QLOG) return;

    for(y=0; y<h; y++){
        for(x=0; x<w; x++){

                src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
            }else if(i>0){
                src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
            }
        }
    }
}

static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
    const int w= b->width;

}

static void correlate_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median, int start_y, int end_y){
    const int w= b->width;
    int x,y;

    IDWTELEM * line=0; // silence silly "could be used without having been initialized" warning
    IDWTELEM * prev;

    if (start_y != 0)

            }else{
                if(y) line[x] += prev[x];
            }
        }
    }
}

static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
    const int w= b->width;
    const int h= b->height;

                for(x=0; x<w; x++){
                    s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
                }
            }
        }
        predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
      }else{
            //ME/MC only
            if(pict->pict_type == I_TYPE){
                for(y=0; y<h; y++){
                    for(x=0; x<w; x++){

                s->mconly_picture.data[plane_index][y*s->mconly_picture.linesize[plane_index] + x]= v;
            }
        }
}

{
    for(level=0; level<s->spatial_decomposition_count; level++){
        for(orientation=level ? 1 : 0; orientation<4; orientation++){
            SubBand *b= &p->band[level][orientation];
            unpack_coeffs(s, b, b->parent, orientation);
        }
    }
}

{
    const int mb_h= s->b_height << s->block_max_depth;
    const int block_size = MB_SIZE >> s->block_max_depth;
    const int block_w    = plane_index ? block_size/2 : block_size;
    int mb_y;
    dwt_compose_t cs[MAX_DECOMPOSITIONS];

        if (!(s->keyframe || s->avctx->debug&512)){
            slice_starty = FFMAX(0, slice_starty - (block_w >> 1));
            slice_h -= (block_w >> 1);
        }

        for(level=0; level<s->spatial_decomposition_count; level++){
            for(orientation=level ? 1 : 0; orientation<4; orientation++){
                SubBand *b= &p->band[level][orientation];
                int start_y;
                int end_y;

                    else
                        decode_subband_slice_buffered(s, b, &s->sb, start_y, end_y, decode_state[level][orientation]);
                }
            }
        }

        for(; yd<slice_h; yd+=4){
            ff_spatial_idwt_buffered_slice(&s->dsp, cs, &s->sb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count, yd);
        }

        if(s->qlog == LOSSLESS_QLOG){
            for(; yq<slice_h && yq<h; yq++){
                IDWTELEM * line = slice_buffer_get_line(&s->sb, yq);
                for(x=0; x<w; x++){

        while(y < end_y)
            slice_buffer_release(&s->sb, y++);
    }

    slice_buffer_flush(&s->sb);
}
    }

    emms_c();

    if(s->last_picture[s->max_ref_frames-1].data[0]){

    memset(s.header_state, 0, sizeof(s.header_state));
    ff_init_range_encoder(&s.c, buffer[0], 256*256);
    ff_init_cabac_states(&s.c, ff_h264_lps_range, ff_h264_mps_state, ff_h264_lps_state, 64);

    for(i=-256; i<256; i++){
        put_symbol(&s.c, s.header_state, i*i*i/3*FFABS(i), 1);
    }
    ff_rac_terminate(&s.c);

    memset(s.header_state, 0, sizeof(s.header_state));
    ff_init_range_decoder(&s.c, buffer[0], 256*256);
    ff_init_cabac_states(&s.c, ff_h264_lps_range, ff_h264_mps_state, ff_h264_lps_state, 64);

    for(i=-256; i<256; i++){
        int j;
        j= get_symbol(&s.c, s.header_state, 1);
        if(j!=i*i*i/3*FFABS(i)) printf("fsck: %d != %d\n", i, j);
    }
#endif
{
int level, orientation, x, y;