comparison snow.c @ 5587:3ae03eacbe9f libavcodec

use 16bit IDWT (a SIMD implementation of it should be >2x faster then with the old 32bit code) disable mmx/sse2 optimizations as they need a rewrite now

5586:f065fc609145

    int stride;
    int width;
    int height;
    int qlog;                                   ///< log(qscale)/log[2^(1/6)]
    DWTELEM *buf;
    int buf_x_offset;
    int buf_y_offset;
    int stride_line; ///< Stride measured in lines, not pixels.
    x_and_coeff * x_coeff;
    struct SubBand *parent;

    int max_ref_frames;
    int ref_frames;
    int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
    uint32_t *ref_scores[MAX_REF_FRAMES];
    DWTELEM *spatial_dwt_buffer;
    int colorspace_type;
    int chroma_h_shift;
    int chroma_v_shift;
    int spatial_scalability;
    int qlog;

    MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to make the motion estimation eventually independent of MpegEncContext, so this will be removed then (FIXME/XXX)
}SnowContext;

typedef struct {
    DWTELEM *b0;
    DWTELEM *b1;
    DWTELEM *b2;
    DWTELEM *b3;
    int y;
} dwt_compose_t;

#define slice_buffer_get_line(slice_buf, line_num) ((slice_buf)->line[line_num] ? (slice_buf)->line[line_num] : slice_buffer_load_line((slice_buf), (line_num)))
//#define slice_buffer_get_line(slice_buf, line_num) (slice_buffer_load_line((slice_buf), (line_num)))

static void iterative_me(SnowContext *s);

static void slice_buffer_init(slice_buffer * buf, int line_count, int max_allocated_lines, int line_width, DWTELEM * base_buffer)
{
    int i;

    buf->base_buffer = base_buffer;
    buf->line_count = line_count;
    buf->line_width = line_width;
    buf->data_count = max_allocated_lines;
    buf->line = (DWTELEM * *) av_mallocz (sizeof(DWTELEM *) * line_count);
    buf->data_stack = (DWTELEM * *) av_malloc (sizeof(DWTELEM *) * max_allocated_lines);

    for (i = 0; i < max_allocated_lines; i++)
    {
      buf->data_stack[i] = (DWTELEM *) av_malloc (sizeof(DWTELEM) * line_width);
    }

    buf->data_stack_top = max_allocated_lines - 1;
}

static DWTELEM * slice_buffer_load_line(slice_buffer * buf, int line)
{
    int offset;
    DWTELEM * buffer;

//  av_log(NULL, AV_LOG_DEBUG, "Cache hit: %d\n", line);

    assert(buf->data_stack_top >= 0);
//  assert(!buf->line[line]);

}

static void slice_buffer_release(slice_buffer * buf, int line)
{
    int offset;
    DWTELEM * buffer;

    assert(line >= 0 && line < buf->line_count);
    assert(buf->line[line]);

    offset = buf->line_width * line;

    if(mirror_right){
        dst[w*dst_step] = LIFT(src[w*src_step], ((mul*2*ref[w*ref_step]+add)>>shift), inverse);
    }
}

#ifndef lift5
static av_always_inline void lift5(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){
    const int mirror_left= !highpass;
    const int mirror_right= (width&1) ^ highpass;
    const int w= (width>>1) - 1 + (highpass & width);

}
#endif

#ifndef liftS
static av_always_inline void liftS(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){
    const int mirror_left= !highpass;
    const int mirror_right= (width&1) ^ highpass;
    const int w= (width>>1) - 1 + (highpass & width);
    int i;

    if(width&1)
        temp[x   ]= b[2*x    ];
    memcpy(b, temp, width*sizeof(int));
}

static void horizontal_composeX(DWTELEM *b, int width){
    DWTELEM temp[width];
    const int width2= width>>1;
    int x;
    const int w2= (width+1)>>1;

    memcpy(temp, b, width*sizeof(int));
    for(x=0; x<width2; x++){
        b[2*x    ]= temp[x   ];
        b[2*x + 1]= temp[x+w2];
    }
    if(width&1)

    inplace_liftV(buffer, width, height, stride, COEFFS2, N2, SHIFT2, LX0, 0);
    inplace_liftV(buffer, width, height, stride, COEFFS3, N3, SHIFT3, LX1, 0);
    inplace_liftV(buffer, width, height, stride, COEFFS4, N4, SHIFT4, LX0, 0);
}

static void spatial_composeX(DWTELEM *buffer, int width, int height, int stride){
    int x, y;

    inplace_liftV(buffer, width, height, stride, COEFFS4, N4, SHIFT4, LX0, 1);
    inplace_liftV(buffer, width, height, stride, COEFFS3, N3, SHIFT3, LX1, 1);
    inplace_liftV(buffer, width, height, stride, COEFFS2, N2, SHIFT2, LX0, 1);

        case DWT_X: spatial_decomposeX  (buffer, width>>level, height>>level, stride<<level); break;
        }
    }
}

static void horizontal_compose53i(DWTELEM *b, int width){
    DWTELEM temp[width];
    const int width2= width>>1;
    const int w2= (width+1)>>1;
    int x;

#if 0

    A3 -= A2;
    A2 += (A1 + A3 + 2)>>2;
    b[width -1] = A3;
    b[width2-1] = A2;
#else
    lift(temp   , b   , b+w2, 1, 1, 1, width,  1, 2, 2, 0, 1);
    lift(temp+w2, b+w2, temp, 1, 1, 1, width, -1, 0, 1, 1, 1);
#endif
    for(x=0; x<width2; x++){
        b[2*x    ]= temp[x   ];
        b[2*x + 1]= temp[x+w2];
    }
    if(width&1)
        b[2*x    ]= temp[x   ];
}

static void vertical_compose53iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
        b1[i] += (b0[i] + b2[i])>>1;
    }
}

static void vertical_compose53iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
        b1[i] -= (b0[i] + b2[i] + 2)>>2;
    }

    cs->b0 = slice_buffer_get_line(sb, mirror(-1-1, height-1) * stride_line);
    cs->b1 = slice_buffer_get_line(sb, mirror(-1  , height-1) * stride_line);
    cs->y = -1;
}

static void spatial_compose53i_init(dwt_compose_t *cs, DWTELEM *buffer, int height, int stride){
    cs->b0 = buffer + mirror(-1-1, height-1)*stride;
    cs->b1 = buffer + mirror(-1  , height-1)*stride;
    cs->y = -1;
}

static void spatial_compose53i_dy_buffered(dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line){
    int y= cs->y;

    DWTELEM *b0= cs->b0;
    DWTELEM *b1= cs->b1;
    DWTELEM *b2= slice_buffer_get_line(sb, mirror(y+1, height-1) * stride_line);
    DWTELEM *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*")}

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

static void spatial_compose53i_dy(dwt_compose_t *cs, DWTELEM *buffer, int width, int height, int stride){
    int y= cs->y;
    DWTELEM *b0= cs->b0;
    DWTELEM *b1= cs->b1;
    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) vertical_compose53iL0(b1, b2, b3, width);
        if(y+0<(unsigned)height) vertical_compose53iH0(b0, b1, b2, width);
STOP_TIMER("vertical_compose53i*")}

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

static void spatial_compose53i(DWTELEM *buffer, int width, int height, int stride){
    dwt_compose_t cs;
    spatial_compose53i_init(&cs, buffer, height, stride);
    while(cs.y <= height)
        spatial_compose53i_dy(&cs, buffer, width, height, stride);
}


void ff_snow_horizontal_compose97i(DWTELEM *b, int width){
    DWTELEM temp[width];
    const int w2= (width+1)>>1;

    lift (temp   , b      , b   +w2, 1, 1, 1, width,  W_DM, W_DO, W_DS, 0, 1);
    lift5(temp+w2, b   +w2, temp   , 1, 1, 1, width,  W_CM, W_CO, W_CS, 1, 1);
    liftS(b      , temp   , temp+w2, 2, 1, 1, width,  W_BM, W_BO, W_BS, 0, 1);
    lift (b+1    , temp+w2, b      , 2, 1, 2, width,  W_AM, W_AO, W_AS, 1, 0);
}

static void vertical_compose97iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
        b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
    }
}

static void vertical_compose97iH1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
#ifdef lift5
        b1[i] -= (W_CM*(b0[i] + b2[i])+W_CO)>>W_CS;

        b1[i] -= (r+W_CO)>>W_CS;
#endif
    }
}

static void vertical_compose97iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
#ifdef liftS
        b1[i] += (W_BM*(b0[i] + b2[i])+W_BO)>>W_BS;

        b1[i] += (W_BM*(b0[i] + b2[i])+4*b1[i]+W_BO)>>W_BS;
#endif
    }
}

static void vertical_compose97iL1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
        b1[i] -= (W_DM*(b0[i] + b2[i])+W_DO)>>W_DS;
    }
}

void ff_snow_vertical_compose97i(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width){
    int i;

    for(i=0; i<width; i++){
#ifndef lift5
        int r;

    cs->b2 = slice_buffer_get_line(sb, mirror(-3+1, height-1) * stride_line);
    cs->b3 = slice_buffer_get_line(sb, mirror(-3+2, height-1) * stride_line);
    cs->y = -3;
}

static void spatial_compose97i_init(dwt_compose_t *cs, DWTELEM *buffer, int height, int stride){
    cs->b0 = buffer + mirror(-3-1, height-1)*stride;
    cs->b1 = buffer + mirror(-3  , height-1)*stride;
    cs->b2 = buffer + mirror(-3+1, height-1)*stride;
    cs->b3 = buffer + mirror(-3+2, height-1)*stride;
    cs->y = -3;
}

static void spatial_compose97i_dy_buffered(DSPContext *dsp, dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line){
    int y = cs->y;

    DWTELEM *b0= cs->b0;
    DWTELEM *b1= cs->b1;
    DWTELEM *b2= cs->b2;
    DWTELEM *b3= cs->b3;
    DWTELEM *b4= slice_buffer_get_line(sb, mirror(y + 3, height - 1) * stride_line);
    DWTELEM *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{

    cs->b2=b4;
    cs->b3=b5;
    cs->y += 2;
}

static void spatial_compose97i_dy(dwt_compose_t *cs, DWTELEM *buffer, int width, int height, int stride){
    int y = cs->y;
    DWTELEM *b0= cs->b0;
    DWTELEM *b1= cs->b1;
    DWTELEM *b2= cs->b2;
    DWTELEM *b3= cs->b3;
    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) 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);

    cs->b2=b4;
    cs->b3=b5;
    cs->y += 2;
}

static void spatial_compose97i(DWTELEM *buffer, int width, int height, int stride){
    dwt_compose_t cs;
    spatial_compose97i_init(&cs, buffer, height, stride);
    while(cs.y <= height)
        spatial_compose97i_dy(&cs, buffer, width, height, stride);
}

          av_log(NULL, AV_LOG_ERROR, "spatial_composeX neither buffered nor slicified yet.\n"); break;
        }
    }
}

static void ff_spatial_idwt_init(dwt_compose_t *cs, DWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count){
    int level;
    for(level=decomposition_count-1; level>=0; level--){
        switch(type){
        case DWT_97: spatial_compose97i_init(cs+level, buffer, height>>level, stride<<level); break;
        case DWT_53: spatial_compose53i_init(cs+level, buffer, height>>level, stride<<level); break;

        case DWT_X: spatial_composeX(buffer, width>>level, height>>level, stride<<level); break;
        }
    }
}

static void ff_spatial_idwt_slice(dwt_compose_t *cs, DWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count, int y){
    const int support = type==1 ? 3 : 5;
    int level;
    if(type==2) return;

    for(level=decomposition_count-1; level>=0; level--){

            }
        }
    }
}

static void ff_spatial_idwt(DWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count){
    if(type==2){
        int level;
        for(level=decomposition_count-1; level>=0; level--)
            spatial_composeX  (buffer, width>>level, height>>level, stride<<level);
    }else{

        for(y=0; y<height; y+=4)
            ff_spatial_idwt_slice(cs, buffer, width, height, stride, type, decomposition_count, y);
    }
}

static int encode_subband_c0run(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){
    const int w= b->width;
    const int h= b->height;
    int x, y;

    if(1){

        }
    }
    return 0;
}

static int encode_subband(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){
//    encode_subband_qtree(s, b, src, parent, stride, orientation);
//    encode_subband_z0run(s, b, src, parent, stride, orientation);
    return encode_subband_c0run(s, b, src, parent, stride, orientation);
//    encode_subband_dzr(s, b, src, parent, stride, orientation);
}

    int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
    int new_index = 0;

    START_TIMER

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

    /* If we are on the second or later slice, restore our index. */

    for(y=start_y; y<h; y++){
        int x = 0;
        int v;
        DWTELEM * line = slice_buffer_get_line(sb, y * b->stride_line + b->buf_y_offset) + b->buf_x_offset;
        memset(line, 0, b->width*sizeof(DWTELEM));
        v = b->x_coeff[new_index].coeff;
        x = b->x_coeff[new_index++].x;
        while(x < w)
        {
            register int t= ( (v>>1)*qmul + qadd)>>QEXPSHIFT;

}

void ff_snow_inner_add_yblock(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
                              int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
    int y, x;
    DWTELEM * dst;
    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);
        const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);

                    +obmc3[x] * block[1][x + y*src_stride]
                    +obmc4[x] * block[0][x + y*src_stride];

            v <<= 8 - LOG2_OBMC_MAX;
            if(FRAC_BITS != 8){
                v += 1<<(7 - FRAC_BITS);
                v >>= 8 - FRAC_BITS;
            }
            if(add){
                v += dst[x + src_x];
                v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;

        }
    }
}

//FIXME name clenup (b_w, block_w, b_width stuff)
static av_always_inline void add_yblock(SnowContext *s, int sliced, slice_buffer *sb, DWTELEM *dst, uint8_t *dst8, const uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int offset_dst, int plane_index){
    const int b_width = s->b_width  << s->block_max_depth;
    const int b_height= s->b_height << s->block_max_depth;
    const int b_stride= b_width;
    BlockNode *lt= &s->block[b_x + b_y*b_stride];
    BlockNode *rt= lt+1;

                    +obmc3[x] * block[1][x + y*src_stride]
                    +obmc4[x] * block[0][x + y*src_stride];

            v <<= 8 - LOG2_OBMC_MAX;
            if(FRAC_BITS != 8){
                v += 1<<(7 - FRAC_BITS);
                v >>= 8 - FRAC_BITS;
            }
            if(add){
                v += dst[x + y*dst_stride];
                v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;

        }
    }
#endif
}

static av_always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * sb, DWTELEM * old_buffer, 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 mb_h= s->b_height << s->block_max_depth;
    int x, y, mb_x;
    int block_size = MB_SIZE >> s->block_max_depth;

        if(add){
            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++)
            {
//                DWTELEM * line = slice_buffer_get_line(sb, y);
                DWTELEM * line = sb->line[y];
                for(x=0; x<w; x++)
                {
//                    int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
                    int v= line[x] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
                    v >>= FRAC_BITS;

            }
        }else{
            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++)
            {
//                DWTELEM * line = slice_buffer_get_line(sb, y);
                DWTELEM * line = sb->line[y];
                for(x=0; x<w; x++)
                {
                    line[x] -= 128 << FRAC_BITS;
//                    buf[x + y*w]-= 128<<FRAC_BITS;
                }

        }

    STOP_TIMER("predict_slice")
}

static av_always_inline void predict_slice(SnowContext *s, DWTELEM *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 mb_h= s->b_height << s->block_max_depth;
    int x, y, mb_x;
    int block_size = MB_SIZE >> s->block_max_depth;

        }

    STOP_TIMER("predict_slice")
}

static av_always_inline void predict_plane(SnowContext *s, DWTELEM *buf, int plane_index, int add){
    const int mb_h= s->b_height << s->block_max_depth;
    int mb_y;
    for(mb_y=0; mb_y<=mb_h; mb_y++)
        predict_slice(s, buf, plane_index, add, mb_y);
}

    const int block_w    = plane_index ? block_size/2 : block_size;
    const uint8_t *obmc  = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth];
    const int obmc_stride= plane_index ? block_size : 2*block_size;
    const int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *src= s-> input_picture.data[plane_index];
    DWTELEM *dst= (DWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
    const int b_stride = s->b_width << s->block_max_depth;
    const int w= p->width;
    const int h= p->height;
    int index= mb_x + mb_y*b_stride;
    BlockNode *b= &s->block[index];

    int ab=0;
    int aa=0;

    b->type|= BLOCK_INTRA;
    b->color[plane_index]= 0;
    memset(dst, 0, obmc_stride*obmc_stride*sizeof(DWTELEM));

    for(i=0; i<4; i++){
        int mb_x2= mb_x + (i &1) - 1;
        int mb_y2= mb_y + (i>>1) - 1;
        int x= block_w*mb_x2 + block_w/2;

    const int block_w    = plane_index ? block_size/2 : block_size;
    const int obmc_stride= plane_index ? block_size : 2*block_size;
    const int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *dst= s->current_picture.data[plane_index];
    uint8_t *src= s->  input_picture.data[plane_index];
    DWTELEM *pred= (DWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
    uint8_t cur[ref_stride*2*MB_SIZE]; //FIXME alignment
    uint8_t tmp[ref_stride*(2*MB_SIZE+5)];
    const int b_stride = s->b_width << s->block_max_depth;
    const int b_height = s->b_height<< s->block_max_depth;
    const int w= p->width;

    pred_block(s, cur, tmp, ref_stride, sx, sy, block_w*2, block_w*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h);

    for(y=y0; y<y1; y++){
        const uint8_t *obmc1= obmc_edged + y*obmc_stride;
        const DWTELEM *pred1 = pred + y*obmc_stride;
        uint8_t *cur1 = cur + y*ref_stride;
        uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
        for(x=x0; x<x1; x++){
            int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
            v = (v + pred1[x]) >> FRAC_BITS;
            if(v&(~255)) v= ~(v>>31);
            dst1[x] = v;
        }
    }

    const uint8_t *obmc  = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth];
    const int obmc_stride= plane_index ? block_size : 2*block_size;
    const int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *dst= s->current_picture.data[plane_index];
    uint8_t *src= s-> input_picture.data[plane_index];
    static const DWTELEM zero_dst[4096]; //FIXME
    const int b_stride = s->b_width << s->block_max_depth;
    const int w= p->width;
    const int h= p->height;
    int distortion= 0;
    int rate= 0;

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

static void quantize(SnowContext *s, SubBand *b, 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) return;

    bias= bias ? 0 : (3*qmul)>>3;
    thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
    thres2= 2*thres1;

                if((unsigned)(i+thres1) > thres2){
                    if(i>=0){
                        i<<= QEXPSHIFT;
                        i/= qmul; //FIXME optimize
                        src[x + y*stride]=  i;
                    }else{
                        i= -i;
                        i<<= QEXPSHIFT;
                        i/= qmul; //FIXME optimize
                        src[x + y*stride]= -i;
                    }
                }else
                    src[x + y*stride]= 0;
            }
        }
    }else{
        for(y=0; y<h; y++){
            for(x=0; x<w; x++){

                if((unsigned)(i+thres1) > thres2){
                    if(i>=0){
                        i<<= QEXPSHIFT;
                        i= (i + bias) / qmul; //FIXME optimize
                        src[x + y*stride]=  i;
                    }else{
                        i= -i;
                        i<<= QEXPSHIFT;
                        i= (i + bias) / qmul; //FIXME optimize
                        src[x + y*stride]= -i;
                    }
                }else
                    src[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, DWTELEM *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));
        DWTELEM * line = slice_buffer_get_line(sb, (y * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;
        for(x=0; x<w; x++){
            int i= line[x];
            if(i<0){
                line[x]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
            }else if(i>0){

    if(w > 200 /*level+1 == s->spatial_decomposition_count*/){
        STOP_TIMER("dquant")
    }
}

static void dequantize(SnowContext *s, SubBand *b, DWTELEM *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;

    if(w > 200 /*level+1 == s->spatial_decomposition_count*/){
        STOP_TIMER("dquant")
    }
}

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

    for(y=h-1; y>=0; y--){

            }
        }
    }
}

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

//    START_TIMER

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

    if (start_y != 0)
        line = slice_buffer_get_line(sb, ((start_y - 1) * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;

    for(y=start_y; y<end_y; y++){

    }

//    STOP_TIMER("correlate")
}

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

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

//    dec += FFMAX(s->chroma_h_shift, s->chroma_v_shift);

    width= s->avctx->width;
    height= s->avctx->height;

    s->spatial_dwt_buffer= av_mallocz(width*height*sizeof(DWTELEM));

    s->mv_scale= (s->avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
    s->block_max_depth= (s->avctx->flags & CODEC_FLAG_4MV) ? 1 : 0;

    for(plane_index=0; plane_index<3; plane_index++){

                }
                if(orientation>1){
                    b->buf += b->stride>>1;
                    b->buf_y_offset = b->stride_line >> 1;
                }

                if(level)
                    b->parent= &s->plane[plane_index].band[level-1][orientation];
                b->x_coeff=av_mallocz(((b->width+1) * b->height+1)*sizeof(x_and_coeff));
            }

    int level, orientation, delta_qlog;

    for(level=0; level<s->spatial_decomposition_count; level++){
        for(orientation=level ? 1 : 0; orientation<4; orientation++){
            SubBand *b= &s->plane[0].band[level][orientation];
            DWTELEM *buf= b->buf;
            const int w= b->width;
            const int h= b->height;
            const int stride= b->stride;
            const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
            const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
            const int qdiv= (1<<16)/qmul;
            int x, y;
            if(orientation==0)
                decorrelate(s, b, buf, stride, 1, 0);
            for(y=0; y<h; y++)
                for(x=0; x<w; x++)
                    coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
            if(orientation==0)
                correlate(s, b, buf, stride, 1, 0);
        }
    }

    /* ugly, ratecontrol just takes a sqrt again */
    coef_sum = (uint64_t)coef_sum * coef_sum >> 16;

    int level, orientation, x, y;

    for(level=0; level<s->spatial_decomposition_count; level++){
        for(orientation=level ? 1 : 0; orientation<4; orientation++){
            SubBand *b= &p->band[level][orientation];
            DWTELEM *buf= b->buf;
            int64_t error=0;

            memset(s->spatial_dwt_buffer, 0, sizeof(int)*width*height);
            buf[b->width/2 + b->height/2*b->stride]= 256*256;
            ff_spatial_idwt(s->spatial_dwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
            for(y=0; y<height; y++){
                for(x=0; x<width; x++){
                    int64_t d= s->spatial_dwt_buffer[x + y*width];
                    error += d*d;
                }
            }

            b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);

    if(!(avctx->flags2 & CODEC_FLAG2_MEMC_ONLY)){
        //FIXME optimize
     if(pict->data[plane_index]) //FIXME gray hack
        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                s->spatial_dwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
            }
        }
        predict_plane(s, s->spatial_dwt_buffer, plane_index, 0);

        if(   plane_index==0
           && pict->pict_type == P_TYPE
           && !(avctx->flags&CODEC_FLAG_PASS2)
           && s->m.me.scene_change_score > s->avctx->scenechange_threshold){

        }

        if(s->qlog == LOSSLESS_QLOG){
            for(y=0; y<h; y++){
                for(x=0; x<w; x++){
                    s->spatial_dwt_buffer[y*w + x]= (s->spatial_dwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
                }
            }
        }else{
            for(y=0; y<h; y++){
                for(x=0; x<w; x++){
                    s->spatial_dwt_buffer[y*w + x]<<=ENCODER_EXTRA_BITS;
                }
            }
        }

        ff_spatial_dwt(s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);

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

                quantize(s, b, b->buf, b->stride, s->qbias);
                if(orientation==0)
                    decorrelate(s, b, b->buf, b->stride, pict->pict_type == P_TYPE, 0);
                encode_subband(s, b, b->buf, b->parent ? b->parent->buf : NULL, b->stride, orientation);
                assert(b->parent==NULL || b->parent->stride == b->stride*2);
                if(orientation==0)
                    correlate(s, b, b->buf, b->stride, 1, 0);
            }
        }
//        av_log(NULL, AV_LOG_DEBUG, "plane:%d bits:%d\n", plane_index, put_bits_count(&s->c.pb) - bits);

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

                dequantize(s, b, b->buf, b->stride);
            }
        }

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

                        s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
                            pict->data[plane_index][y*pict->linesize[plane_index] + x];
                    }
                }
            }else{
                memset(s->spatial_dwt_buffer, 0, sizeof(DWTELEM)*w*h);
                predict_plane(s, s->spatial_dwt_buffer, plane_index, 1);
            }
      }
        if(s->avctx->flags&CODEC_FLAG_PSNR){
            int64_t error= 0;

static void common_end(SnowContext *s){
    int plane_index, level, orientation, i;

    av_freep(&s->spatial_dwt_buffer);

    av_freep(&s->m.me.scratchpad);
    av_freep(&s->m.me.map);
    av_freep(&s->m.me.score_map);
    av_freep(&s->m.obmc_scratchpad);

    avctx->pix_fmt= PIX_FMT_YUV420P;

    common_init(avctx);

    block_size = MB_SIZE >> s->block_max_depth;
    slice_buffer_init(&s->sb, s->plane[0].height, (block_size) + (s->spatial_decomposition_count * (s->spatial_decomposition_count + 3)) + 1, s->plane[0].width, s->spatial_dwt_buffer);

    return 0;
}

static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8_t *buf, int buf_size){

        int x, y;
        int decode_state[MAX_DECOMPOSITIONS][4][1]; /* Stored state info for unpack_coeffs. 1 variable per instance. */

if(s->avctx->debug&2048){
        memset(s->spatial_dwt_buffer, 0, sizeof(DWTELEM)*w*h);
        predict_plane(s, s->spatial_dwt_buffer, plane_index, 1);

        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                int v= s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x];
                s->mconly_picture.data[plane_index][y*s->mconly_picture.linesize[plane_index] + x]= v;

                    if (orientation == 0){
                        SubBand * correlate_band = &p->band[0][0];
                        int correlate_end_y = FFMIN(b->height, end_y + 1);
                        int correlate_start_y = FFMIN(b->height, (start_y ? start_y + 1 : 0));
                        decode_subband_slice_buffered(s, correlate_band, &s->sb, correlate_start_y, correlate_end_y, decode_state[0][0]);
                        correlate_slice_buffered(s, &s->sb, correlate_band, correlate_band->buf, correlate_band->stride, 1, 0, correlate_start_y, correlate_end_y);
                        dequantize_slice_buffered(s, &s->sb, correlate_band, correlate_band->buf, correlate_band->stride, start_y, end_y);
                    }
                    else
                        decode_subband_slice_buffered(s, b, &s->sb, start_y, end_y, decode_state[level][orientation]);
                }
            }

    STOP_TIMER("idwt slice");}


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

        predict_slice_buffered(s, &s->sb, s->spatial_dwt_buffer, plane_index, 1, mb_y);

        y = FFMIN(p->height, slice_starty);
        end_y = FFMIN(p->height, slice_h);
        while(y < end_y)
            slice_buffer_release(&s->sb, y++);

5587:3ae03eacbe9f

    int stride;
    int width;
    int height;
    int qlog;                                   ///< log(qscale)/log[2^(1/6)]
    DWTELEM *buf;
    IDWTELEM *ibuf;
    int buf_x_offset;
    int buf_y_offset;
    int stride_line; ///< Stride measured in lines, not pixels.
    x_and_coeff * x_coeff;
    struct SubBand *parent;

    int max_ref_frames;
    int ref_frames;
    int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
    uint32_t *ref_scores[MAX_REF_FRAMES];
    DWTELEM *spatial_dwt_buffer;
    IDWTELEM *spatial_idwt_buffer;
    int colorspace_type;
    int chroma_h_shift;
    int chroma_v_shift;
    int spatial_scalability;
    int qlog;

    MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to make the motion estimation eventually independent of MpegEncContext, so this will be removed then (FIXME/XXX)
}SnowContext;

typedef struct {
    IDWTELEM *b0;
    IDWTELEM *b1;
    IDWTELEM *b2;
    IDWTELEM *b3;
    int y;
} dwt_compose_t;

#define slice_buffer_get_line(slice_buf, line_num) ((slice_buf)->line[line_num] ? (slice_buf)->line[line_num] : slice_buffer_load_line((slice_buf), (line_num)))
//#define slice_buffer_get_line(slice_buf, line_num) (slice_buffer_load_line((slice_buf), (line_num)))

static void iterative_me(SnowContext *s);

static void slice_buffer_init(slice_buffer * buf, int line_count, int max_allocated_lines, int line_width, IDWTELEM * base_buffer)
{
    int i;

    buf->base_buffer = base_buffer;
    buf->line_count = line_count;
    buf->line_width = line_width;
    buf->data_count = max_allocated_lines;
    buf->line = av_mallocz (sizeof(IDWTELEM *) * line_count);
    buf->data_stack = av_malloc (sizeof(IDWTELEM *) * max_allocated_lines);

    for (i = 0; i < max_allocated_lines; i++)
    {
        buf->data_stack[i] = av_malloc (sizeof(IDWTELEM) * line_width);
    }

    buf->data_stack_top = max_allocated_lines - 1;
}

static IDWTELEM * slice_buffer_load_line(slice_buffer * buf, int line)
{
    int offset;
    IDWTELEM * buffer;

//  av_log(NULL, AV_LOG_DEBUG, "Cache hit: %d\n", line);

    assert(buf->data_stack_top >= 0);
//  assert(!buf->line[line]);

}

static void slice_buffer_release(slice_buffer * buf, int line)
{
    int offset;
    IDWTELEM * buffer;

    assert(line >= 0 && line < buf->line_count);
    assert(buf->line[line]);

    offset = buf->line_width * line;

    if(mirror_right){
        dst[w*dst_step] = LIFT(src[w*src_step], ((mul*2*ref[w*ref_step]+add)>>shift), inverse);
    }
}

static av_always_inline void inv_lift(IDWTELEM *dst, IDWTELEM *src, IDWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){
    const int mirror_left= !highpass;
    const int mirror_right= (width&1) ^ highpass;
    const int w= (width>>1) - 1 + (highpass & width);
    int i;

#define LIFT(src, ref, inv) ((src) + ((inv) ? - (ref) : + (ref)))
    if(mirror_left){
        dst[0] = LIFT(src[0], ((mul*2*ref[0]+add)>>shift), inverse);
        dst += dst_step;
        src += src_step;
    }

    for(i=0; i<w; i++){
        dst[i*dst_step] = LIFT(src[i*src_step], ((mul*(ref[i*ref_step] + ref[(i+1)*ref_step])+add)>>shift), inverse);
    }

    if(mirror_right){
        dst[w*dst_step] = LIFT(src[w*src_step], ((mul*2*ref[w*ref_step]+add)>>shift), inverse);
    }
}


#ifndef lift5
static av_always_inline void lift5(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){
    const int mirror_left= !highpass;
    const int mirror_right= (width&1) ^ highpass;
    const int w= (width>>1) - 1 + (highpass & width);

}
#endif

#ifndef liftS
static av_always_inline void liftS(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){
    const int mirror_left= !highpass;
    const int mirror_right= (width&1) ^ highpass;
    const int w= (width>>1) - 1 + (highpass & width);
    int i;

    assert(shift == 4);
#define LIFTS(src, ref, inv) ((inv) ? (src) + (((ref) + 4*(src))>>shift): -((-16*(src) + (ref) + add/4 + 1 + (5<<25))/(5*4) - (1<<23)))
    if(mirror_left){
        dst[0] = LIFTS(src[0], mul*2*ref[0]+add, inverse);
        dst += dst_step;
        src += src_step;
    }

    for(i=0; i<w; i++){
        dst[i*dst_step] = LIFTS(src[i*src_step], mul*(ref[i*ref_step] + ref[(i+1)*ref_step])+add, inverse);
    }

    if(mirror_right){
        dst[w*dst_step] = LIFTS(src[w*src_step], mul*2*ref[w*ref_step]+add, inverse);
    }
}
static av_always_inline void inv_liftS(IDWTELEM *dst, IDWTELEM *src, IDWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){
    const int mirror_left= !highpass;
    const int mirror_right= (width&1) ^ highpass;
    const int w= (width>>1) - 1 + (highpass & width);
    int i;

    if(width&1)
        temp[x   ]= b[2*x    ];
    memcpy(b, temp, width*sizeof(int));
}

static void horizontal_composeX(IDWTELEM *b, int width){
    IDWTELEM temp[width];
    const int width2= width>>1;
    int x;
    const int w2= (width+1)>>1;

    memcpy(temp, b, width*sizeof(IDWTELEM));
    for(x=0; x<width2; x++){
        b[2*x    ]= temp[x   ];
        b[2*x + 1]= temp[x+w2];
    }
    if(width&1)

    inplace_liftV(buffer, width, height, stride, COEFFS2, N2, SHIFT2, LX0, 0);
    inplace_liftV(buffer, width, height, stride, COEFFS3, N3, SHIFT3, LX1, 0);
    inplace_liftV(buffer, width, height, stride, COEFFS4, N4, SHIFT4, LX0, 0);
}

static void spatial_composeX(IDWTELEM *buffer, int width, int height, int stride){
    int x, y;

    inplace_liftV(buffer, width, height, stride, COEFFS4, N4, SHIFT4, LX0, 1);
    inplace_liftV(buffer, width, height, stride, COEFFS3, N3, SHIFT3, LX1, 1);
    inplace_liftV(buffer, width, height, stride, COEFFS2, N2, SHIFT2, LX0, 1);

        case DWT_X: spatial_decomposeX  (buffer, width>>level, height>>level, stride<<level); break;
        }
    }
}

static void horizontal_compose53i(IDWTELEM *b, int width){
    IDWTELEM temp[width];
    const int width2= width>>1;
    const int w2= (width+1)>>1;
    int x;

#if 0

    A3 -= A2;
    A2 += (A1 + A3 + 2)>>2;
    b[width -1] = A3;
    b[width2-1] = A2;
#else
    inv_lift(temp   , b   , b+w2, 1, 1, 1, width,  1, 2, 2, 0, 1);
    inv_lift(temp+w2, b+w2, temp, 1, 1, 1, width, -1, 0, 1, 1, 1);
#endif
    for(x=0; x<width2; x++){
        b[2*x    ]= temp[x   ];
        b[2*x + 1]= temp[x+w2];
    }
    if(width&1)
        b[2*x    ]= temp[x   ];
}

static void vertical_compose53iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
        b1[i] += (b0[i] + b2[i])>>1;
    }
}

static void vertical_compose53iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
        b1[i] -= (b0[i] + b2[i] + 2)>>2;
    }

    cs->b0 = slice_buffer_get_line(sb, mirror(-1-1, height-1) * stride_line);
    cs->b1 = slice_buffer_get_line(sb, mirror(-1  , height-1) * stride_line);
    cs->y = -1;
}

static void spatial_compose53i_init(dwt_compose_t *cs, IDWTELEM *buffer, int height, int stride){
    cs->b0 = buffer + mirror(-1-1, height-1)*stride;
    cs->b1 = buffer + mirror(-1  , height-1)*stride;
    cs->y = -1;
}

static void spatial_compose53i_dy_buffered(dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line){
    int y= cs->y;

    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*")}

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

static void spatial_compose53i_dy(dwt_compose_t *cs, IDWTELEM *buffer, int width, int height, int stride){
    int y= cs->y;
    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*")}

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

static void spatial_compose53i(IDWTELEM *buffer, int width, int height, int stride){
    dwt_compose_t cs;
    spatial_compose53i_init(&cs, buffer, height, stride);
    while(cs.y <= height)
        spatial_compose53i_dy(&cs, buffer, width, height, stride);
}


void ff_snow_horizontal_compose97i(IDWTELEM *b, int width){
    IDWTELEM temp[width];
    const int w2= (width+1)>>1;

    inv_lift (temp   , b      , b   +w2, 1, 1, 1, width,  W_DM, W_DO, W_DS, 0, 1);
    inv_lift5(temp+w2, b   +w2, temp   , 1, 1, 1, width,  W_CM, W_CO, W_CS, 1, 1);
    inv_liftS(b      , temp   , temp+w2, 2, 1, 1, width,  W_BM, W_BO, W_BS, 0, 1);
    inv_lift (b+1    , temp+w2, b      , 2, 1, 2, width,  W_AM, W_AO, W_AS, 1, 0);
}

static void vertical_compose97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
        b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
    }
}

static void vertical_compose97iH1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
#ifdef lift5
        b1[i] -= (W_CM*(b0[i] + b2[i])+W_CO)>>W_CS;

        b1[i] -= (r+W_CO)>>W_CS;
#endif
    }
}

static void vertical_compose97iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
#ifdef liftS
        b1[i] += (W_BM*(b0[i] + b2[i])+W_BO)>>W_BS;

        b1[i] += (W_BM*(b0[i] + b2[i])+4*b1[i]+W_BO)>>W_BS;
#endif
    }
}

static void vertical_compose97iL1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width){
    int i;

    for(i=0; i<width; i++){
        b1[i] -= (W_DM*(b0[i] + b2[i])+W_DO)>>W_DS;
    }
}

void ff_snow_vertical_compose97i(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
    int i;

    for(i=0; i<width; i++){
#ifndef lift5
        int r;

    cs->b2 = slice_buffer_get_line(sb, mirror(-3+1, height-1) * stride_line);
    cs->b3 = slice_buffer_get_line(sb, mirror(-3+2, height-1) * stride_line);
    cs->y = -3;
}

static void spatial_compose97i_init(dwt_compose_t *cs, IDWTELEM *buffer, int height, int stride){
    cs->b0 = buffer + mirror(-3-1, height-1)*stride;
    cs->b1 = buffer + mirror(-3  , height-1)*stride;
    cs->b2 = buffer + mirror(-3+1, height-1)*stride;
    cs->b3 = buffer + mirror(-3+2, height-1)*stride;
    cs->y = -3;
}

static void spatial_compose97i_dy_buffered(DSPContext *dsp, dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line){
    int y = cs->y;

    IDWTELEM *b0= cs->b0;
    IDWTELEM *b1= cs->b1;
    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{

    cs->b2=b4;
    cs->b3=b5;
    cs->y += 2;
}

static void spatial_compose97i_dy(dwt_compose_t *cs, IDWTELEM *buffer, int width, int height, int stride){
    int y = cs->y;
    IDWTELEM *b0= cs->b0;
    IDWTELEM *b1= cs->b1;
    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);

    cs->b2=b4;
    cs->b3=b5;
    cs->y += 2;
}

static void spatial_compose97i(IDWTELEM *buffer, int width, int height, int stride){
    dwt_compose_t cs;
    spatial_compose97i_init(&cs, buffer, height, stride);
    while(cs.y <= height)
        spatial_compose97i_dy(&cs, buffer, width, height, stride);
}

          av_log(NULL, AV_LOG_ERROR, "spatial_composeX neither buffered nor slicified yet.\n"); break;
        }
    }
}

static void ff_spatial_idwt_init(dwt_compose_t *cs, IDWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count){
    int level;
    for(level=decomposition_count-1; level>=0; level--){
        switch(type){
        case DWT_97: spatial_compose97i_init(cs+level, buffer, height>>level, stride<<level); break;
        case DWT_53: spatial_compose53i_init(cs+level, buffer, height>>level, stride<<level); break;

        case DWT_X: spatial_composeX(buffer, width>>level, height>>level, stride<<level); break;
        }
    }
}

static void ff_spatial_idwt_slice(dwt_compose_t *cs, IDWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count, int y){
    const int support = type==1 ? 3 : 5;
    int level;
    if(type==2) return;

    for(level=decomposition_count-1; level>=0; level--){

            }
        }
    }
}

static void ff_spatial_idwt(IDWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count){
    if(type==2){
        int level;
        for(level=decomposition_count-1; level>=0; level--)
            spatial_composeX  (buffer, width>>level, height>>level, stride<<level);
    }else{

        for(y=0; y<height; y+=4)
            ff_spatial_idwt_slice(cs, buffer, width, height, stride, type, decomposition_count, y);
    }
}

static int encode_subband_c0run(SnowContext *s, SubBand *b, IDWTELEM *src, IDWTELEM *parent, int stride, int orientation){
    const int w= b->width;
    const int h= b->height;
    int x, y;

    if(1){

        }
    }
    return 0;
}

static int encode_subband(SnowContext *s, SubBand *b, IDWTELEM *src, IDWTELEM *parent, int stride, int orientation){
//    encode_subband_qtree(s, b, src, parent, stride, orientation);
//    encode_subband_z0run(s, b, src, parent, stride, orientation);
    return encode_subband_c0run(s, b, src, parent, stride, orientation);
//    encode_subband_dzr(s, b, src, parent, stride, orientation);
}

    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;
    }

    /* If we are on the second or later slice, restore our index. */

    for(y=start_y; y<h; y++){
        int x = 0;
        int v;
        IDWTELEM * line = slice_buffer_get_line(sb, y * b->stride_line + b->buf_y_offset) + b->buf_x_offset;
        memset(line, 0, b->width*sizeof(IDWTELEM));
        v = b->x_coeff[new_index].coeff;
        x = b->x_coeff[new_index++].x;
        while(x < w)
        {
            register int t= ( (v>>1)*qmul + qadd)>>QEXPSHIFT;

}

void ff_snow_inner_add_yblock(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
                              int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
    int y, x;
    IDWTELEM * dst;
    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);
        const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);

                    +obmc3[x] * block[1][x + y*src_stride]
                    +obmc4[x] * block[0][x + y*src_stride];

            v <<= 8 - LOG2_OBMC_MAX;
            if(FRAC_BITS != 8){
                v >>= 8 - FRAC_BITS;
            }
            if(add){
                v += dst[x + src_x];
                v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;

        }
    }
}

//FIXME name clenup (b_w, block_w, b_width stuff)
static av_always_inline void add_yblock(SnowContext *s, int sliced, slice_buffer *sb, IDWTELEM *dst, uint8_t *dst8, const uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int offset_dst, int plane_index){
    const int b_width = s->b_width  << s->block_max_depth;
    const int b_height= s->b_height << s->block_max_depth;
    const int b_stride= b_width;
    BlockNode *lt= &s->block[b_x + b_y*b_stride];
    BlockNode *rt= lt+1;

                    +obmc3[x] * block[1][x + y*src_stride]
                    +obmc4[x] * block[0][x + y*src_stride];

            v <<= 8 - LOG2_OBMC_MAX;
            if(FRAC_BITS != 8){
                v >>= 8 - FRAC_BITS;
            }
            if(add){
                v += dst[x + y*dst_stride];
                v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;

        }
    }
#endif
}

static av_always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * sb, IDWTELEM * old_buffer, 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 mb_h= s->b_height << s->block_max_depth;
    int x, y, mb_x;
    int block_size = MB_SIZE >> s->block_max_depth;

        if(add){
            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++)
            {
//                DWTELEM * line = slice_buffer_get_line(sb, y);
                IDWTELEM * line = sb->line[y];
                for(x=0; x<w; x++)
                {
//                    int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
                    int v= line[x] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
                    v >>= FRAC_BITS;

            }
        }else{
            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++)
            {
//                DWTELEM * line = slice_buffer_get_line(sb, y);
                IDWTELEM * line = sb->line[y];
                for(x=0; x<w; x++)
                {
                    line[x] -= 128 << FRAC_BITS;
//                    buf[x + y*w]-= 128<<FRAC_BITS;
                }

        }

    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 mb_h= s->b_height << s->block_max_depth;
    int x, y, mb_x;
    int block_size = MB_SIZE >> s->block_max_depth;

        }

    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;
    for(mb_y=0; mb_y<=mb_h; mb_y++)
        predict_slice(s, buf, plane_index, add, mb_y);
}

    const int block_w    = plane_index ? block_size/2 : block_size;
    const uint8_t *obmc  = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth];
    const int obmc_stride= plane_index ? block_size : 2*block_size;
    const int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *src= s-> input_picture.data[plane_index];
    IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
    const int b_stride = s->b_width << s->block_max_depth;
    const int w= p->width;
    const int h= p->height;
    int index= mb_x + mb_y*b_stride;
    BlockNode *b= &s->block[index];

    int ab=0;
    int aa=0;

    b->type|= BLOCK_INTRA;
    b->color[plane_index]= 0;
    memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));

    for(i=0; i<4; i++){
        int mb_x2= mb_x + (i &1) - 1;
        int mb_y2= mb_y + (i>>1) - 1;
        int x= block_w*mb_x2 + block_w/2;

    const int block_w    = plane_index ? block_size/2 : block_size;
    const int obmc_stride= plane_index ? block_size : 2*block_size;
    const int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *dst= s->current_picture.data[plane_index];
    uint8_t *src= s->  input_picture.data[plane_index];
    IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
    uint8_t cur[ref_stride*2*MB_SIZE]; //FIXME alignment
    uint8_t tmp[ref_stride*(2*MB_SIZE+5)];
    const int b_stride = s->b_width << s->block_max_depth;
    const int b_height = s->b_height<< s->block_max_depth;
    const int w= p->width;

    pred_block(s, cur, tmp, ref_stride, sx, sy, block_w*2, block_w*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h);

    for(y=y0; y<y1; y++){
        const uint8_t *obmc1= obmc_edged + y*obmc_stride;
        const IDWTELEM *pred1 = pred + y*obmc_stride;
        uint8_t *cur1 = cur + y*ref_stride;
        uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
        for(x=x0; x<x1; x++){
#if FRAC_BITS >= LOG2_OBMC_MAX
            int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
#else
            int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
#endif
            v = (v + pred1[x]) >> FRAC_BITS;
            if(v&(~255)) v= ~(v>>31);
            dst1[x] = v;
        }
    }

    const uint8_t *obmc  = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth];
    const int obmc_stride= plane_index ? block_size : 2*block_size;
    const int ref_stride= s->current_picture.linesize[plane_index];
    uint8_t *dst= s->current_picture.data[plane_index];
    uint8_t *src= s-> input_picture.data[plane_index];
    static const IDWTELEM zero_dst[4096]; //FIXME
    const int b_stride = s->b_width << s->block_max_depth;
    const int w= p->width;
    const int h= p->height;
    int distortion= 0;
    int rate= 0;

        }
        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];
        return;
    }

    bias= bias ? 0 : (3*qmul)>>3;
    thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
    thres2= 2*thres1;

                if((unsigned)(i+thres1) > thres2){
                    if(i>=0){
                        i<<= QEXPSHIFT;
                        i/= qmul; //FIXME optimize
                        dst[x + y*stride]=  i;
                    }else{
                        i= -i;
                        i<<= QEXPSHIFT;
                        i/= qmul; //FIXME optimize
                        dst[x + y*stride]= -i;
                    }
                }else
                    dst[x + y*stride]= 0;
            }
        }
    }else{
        for(y=0; y<h; y++){
            for(x=0; x<w; x++){

                if((unsigned)(i+thres1) > thres2){
                    if(i>=0){
                        i<<= QEXPSHIFT;
                        i= (i + bias) / qmul; //FIXME optimize
                        dst[x + y*stride]=  i;
                    }else{
                        i= -i;
                        i<<= QEXPSHIFT;
                        i= (i + bias) / qmul; //FIXME optimize
                        dst[x + y*stride]= -i;
                    }
                }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;

    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));
        IDWTELEM * line = slice_buffer_get_line(sb, (y * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;
        for(x=0; x<w; x++){
            int i= line[x];
            if(i<0){
                line[x]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
            }else if(i>0){

    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;

    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;
    const int h= b->height;
    int x,y;

    for(y=h-1; y>=0; y--){

            }
        }
    }
}

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)
        line = slice_buffer_get_line(sb, ((start_y - 1) * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;

    for(y=start_y; y<end_y; y++){

    }

//    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;
    int x,y;

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

//    dec += FFMAX(s->chroma_h_shift, s->chroma_v_shift);

    width= s->avctx->width;
    height= s->avctx->height;

    s->spatial_idwt_buffer= av_mallocz(width*height*sizeof(IDWTELEM));
    s->spatial_dwt_buffer= av_mallocz(width*height*sizeof(DWTELEM)); //FIXME this doesnt belong here

    s->mv_scale= (s->avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
    s->block_max_depth= (s->avctx->flags & CODEC_FLAG_4MV) ? 1 : 0;

    for(plane_index=0; plane_index<3; plane_index++){

                }
                if(orientation>1){
                    b->buf += b->stride>>1;
                    b->buf_y_offset = b->stride_line >> 1;
                }
                b->ibuf= s->spatial_idwt_buffer + (b->buf - s->spatial_dwt_buffer);

                if(level)
                    b->parent= &s->plane[plane_index].band[level-1][orientation];
                b->x_coeff=av_mallocz(((b->width+1) * b->height+1)*sizeof(x_and_coeff));
            }

    int level, orientation, delta_qlog;

    for(level=0; level<s->spatial_decomposition_count; level++){
        for(orientation=level ? 1 : 0; orientation<4; orientation++){
            SubBand *b= &s->plane[0].band[level][orientation];
            IDWTELEM *buf= b->ibuf;
            const int w= b->width;
            const int h= b->height;
            const int stride= b->stride;
            const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
            const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
            const int qdiv= (1<<16)/qmul;
            int x, y;
            //FIXME this is ugly
            for(y=0; y<h; y++)
                for(x=0; x<w; x++)
                    buf[x+y*stride]= b->buf[x+y*stride];
            if(orientation==0)
                decorrelate(s, b, buf, stride, 1, 0);
            for(y=0; y<h; y++)
                for(x=0; x<w; x++)
                    coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
        }
    }

    /* ugly, ratecontrol just takes a sqrt again */
    coef_sum = (uint64_t)coef_sum * coef_sum >> 16;

    int level, orientation, x, y;

    for(level=0; level<s->spatial_decomposition_count; level++){
        for(orientation=level ? 1 : 0; orientation<4; orientation++){
            SubBand *b= &p->band[level][orientation];
            IDWTELEM *ibuf= b->ibuf;
            int64_t error=0;

            memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
            ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
            ff_spatial_idwt(s->spatial_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
            for(y=0; y<height; y++){
                for(x=0; x<width; x++){
                    int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
                    error += d*d;
                }
            }

            b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);

    if(!(avctx->flags2 & CODEC_FLAG2_MEMC_ONLY)){
        //FIXME optimize
     if(pict->data[plane_index]) //FIXME gray hack
        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
            }
        }
        predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);

        if(   plane_index==0
           && pict->pict_type == P_TYPE
           && !(avctx->flags&CODEC_FLAG_PASS2)
           && s->m.me.scene_change_score > s->avctx->scenechange_threshold){

        }

        if(s->qlog == LOSSLESS_QLOG){
            for(y=0; y<h; y++){
                for(x=0; x<w; x++){
                    s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
                }
            }
        }else{
            for(y=0; y<h; y++){
                for(x=0; x<w; x++){
                    s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
                }
            }
        }

        ff_spatial_dwt(s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);

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

                quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
                if(orientation==0)
                    decorrelate(s, b, b->ibuf, b->stride, pict->pict_type == P_TYPE, 0);
                encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
                assert(b->parent==NULL || b->parent->stride == b->stride*2);
                if(orientation==0)
                    correlate(s, b, b->ibuf, b->stride, 1, 0);
            }
        }
//        av_log(NULL, AV_LOG_DEBUG, "plane:%d bits:%d\n", plane_index, put_bits_count(&s->c.pb) - bits);

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

                dequantize(s, b, b->ibuf, b->stride);
            }
        }

        ff_spatial_idwt(s->spatial_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
        if(s->qlog == LOSSLESS_QLOG){
            for(y=0; y<h; y++){
                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++){

                        s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
                            pict->data[plane_index][y*pict->linesize[plane_index] + x];
                    }
                }
            }else{
                memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
                predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
            }
      }
        if(s->avctx->flags&CODEC_FLAG_PSNR){
            int64_t error= 0;

static void common_end(SnowContext *s){
    int plane_index, level, orientation, i;

    av_freep(&s->spatial_dwt_buffer);
    av_freep(&s->spatial_idwt_buffer);

    av_freep(&s->m.me.scratchpad);
    av_freep(&s->m.me.map);
    av_freep(&s->m.me.score_map);
    av_freep(&s->m.obmc_scratchpad);

    avctx->pix_fmt= PIX_FMT_YUV420P;

    common_init(avctx);

    block_size = MB_SIZE >> s->block_max_depth;
    slice_buffer_init(&s->sb, s->plane[0].height, (block_size) + (s->spatial_decomposition_count * (s->spatial_decomposition_count + 3)) + 1, s->plane[0].width, s->spatial_idwt_buffer);

    return 0;
}

static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8_t *buf, int buf_size){

        int x, y;
        int decode_state[MAX_DECOMPOSITIONS][4][1]; /* Stored state info for unpack_coeffs. 1 variable per instance. */

if(s->avctx->debug&2048){
        memset(s->spatial_dwt_buffer, 0, sizeof(DWTELEM)*w*h);
        predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);

        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                int v= s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x];
                s->mconly_picture.data[plane_index][y*s->mconly_picture.linesize[plane_index] + x]= v;

                    if (orientation == 0){
                        SubBand * correlate_band = &p->band[0][0];
                        int correlate_end_y = FFMIN(b->height, end_y + 1);
                        int correlate_start_y = FFMIN(b->height, (start_y ? start_y + 1 : 0));
                        decode_subband_slice_buffered(s, correlate_band, &s->sb, correlate_start_y, correlate_end_y, decode_state[0][0]);
                        correlate_slice_buffered(s, &s->sb, correlate_band, correlate_band->ibuf, correlate_band->stride, 1, 0, correlate_start_y, correlate_end_y);
                        dequantize_slice_buffered(s, &s->sb, correlate_band, correlate_band->ibuf, correlate_band->stride, start_y, end_y);
                    }
                    else
                        decode_subband_slice_buffered(s, b, &s->sb, start_y, end_y, decode_state[level][orientation]);
                }
            }

    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++){
                    line[x] <<= FRAC_BITS;
                }
            }
        }

        predict_slice_buffered(s, &s->sb, s->spatial_idwt_buffer, plane_index, 1, mb_y);

        y = FFMIN(p->height, slice_starty);
        end_y = FFMIN(p->height, slice_h);
        while(y < end_y)
            slice_buffer_release(&s->sb, y++);