Merge remote branch 'qatar/master'
[ffmpeg.git] / libavcodec / snow.c
index 44cfd44..4ee1722 100644 (file)
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
+#include "libavutil/intmath.h"
 #include "avcodec.h"
 #include "dsputil.h"
+#include "dwt.h"
 #include "snow.h"
 
 #include "rangecoder.h"
 #include "mathops.h"
 
 #include "mpegvideo.h"
+#include "h263.h"
 
 #undef NDEBUG
 #include <assert.h>
@@ -434,11 +437,11 @@ typedef struct Plane{
 }Plane;
 
 typedef struct SnowContext{
-//    MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to eventually make the motion estimation independent of MpegEncContext, so this will be removed then (FIXME/XXX)
 
     AVCodecContext *avctx;
     RangeCoder c;
     DSPContext dsp;
+    DWTContext dwt;
     AVFrame new_picture;
     AVFrame input_picture;              ///< new_picture with the internal linesizes
     AVFrame current_picture;
@@ -493,91 +496,6 @@ typedef struct SnowContext{
     uint8_t *scratchbuf;
 }SnowContext;
 
-typedef struct {
-    IDWTELEM *b0;
-    IDWTELEM *b1;
-    IDWTELEM *b2;
-    IDWTELEM *b3;
-    int y;
-} DWTCompose;
-
-#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;
-
-    assert(buf->data_stack_top >= 0);
-//  assert(!buf->line[line]);
-    if (buf->line[line])
-        return buf->line[line];
-
-    offset = buf->line_width * line;
-    buffer = buf->data_stack[buf->data_stack_top];
-    buf->data_stack_top--;
-    buf->line[line] = buffer;
-
-    return buffer;
-}
-
-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;
-    buffer = buf->line[line];
-    buf->data_stack_top++;
-    buf->data_stack[buf->data_stack_top] = buffer;
-    buf->line[line] = NULL;
-}
-
-static void slice_buffer_flush(slice_buffer * buf)
-{
-    int i;
-    for(i = 0; i < buf->line_count; i++){
-        if (buf->line[i])
-            slice_buffer_release(buf, i);
-    }
-}
-
-static void slice_buffer_destroy(slice_buffer * buf)
-{
-    int i;
-    slice_buffer_flush(buf);
-
-    for(i = buf->data_count - 1; i >= 0; i--){
-        av_freep(&buf->data_stack[i]);
-    }
-    av_freep(&buf->data_stack);
-    av_freep(&buf->line);
-}
-
 #ifdef __sgi
 // Avoid a name clash on SGI IRIX
 #undef qexp
@@ -585,14 +503,6 @@ static void slice_buffer_destroy(slice_buffer * buf)
 #define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
 static uint8_t qexp[QROOT];
 
-static inline int mirror(int v, int m){
-    while((unsigned)v > (unsigned)m){
-        v=-v;
-        if(v<0) v+= 2*m;
-    }
-    return v;
-}
-
 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
     int i;
 
@@ -639,14 +549,14 @@ static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signe
             for(i=0; i<e; i++){
                 put_rac(c, state+1+FFMIN(i,9), 1);  //1..10
             }
-            put_rac(c, state+1+FFMIN(i,9), 0);
+            put_rac(c, state+1+9, 0);
 
             for(i=e-1; i>=0; i--){
                 put_rac(c, state+22+FFMIN(i,9), (a>>i)&1); //22..31
             }
 
             if(is_signed)
-                put_rac(c, state+11 + FFMIN(e,10), v < 0); //11..21
+                put_rac(c, state+11 + 10, v < 0); //11..21
         }
 #endif /* 1 */
     }else{
@@ -669,10 +579,8 @@ static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
             a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
         }
 
-        if(is_signed && get_rac(c, state+11 + FFMIN(e,10))) //11..21
-            return -a;
-        else
-            return a;
+        e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
+        return (a^e)-e;
     }
 }
 
@@ -716,3445 +624,3074 @@ static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
     return v;
 }
 
-static av_always_inline void
-lift(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;
-
-#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;
-    }
+static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
+    const int w= b->width;
+    const int h= b->height;
+    int x,y;
 
-    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);
-    }
+    int run, runs;
+    x_and_coeff *xc= b->x_coeff;
+    x_and_coeff *prev_xc= NULL;
+    x_and_coeff *prev2_xc= xc;
+    x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
+    x_and_coeff *prev_parent_xc= parent_xc;
 
-    if(mirror_right){
-        dst[w*dst_step] =
-            LIFT(src[w*src_step],
-                 ((mul*2*ref[w*ref_step]+add)>>shift),
-                 inverse);
-    }
-}
+    runs= get_symbol2(&s->c, b->state[30], 0);
+    if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
+    else           run= INT_MAX;
 
-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;
+    for(y=0; y<h; y++){
+        int v=0;
+        int lt=0, t=0, rt=0;
 
-#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;
-    }
+        if(y && prev_xc->x == 0){
+            rt= prev_xc->coeff;
+        }
+        for(x=0; x<w; x++){
+            int p=0;
+            const int l= v;
 
-    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);
-    }
+            lt= t; t= rt;
 
-    if(mirror_right){
-        dst[w*dst_step] =
-            LIFT(src[w*src_step],
-                 ((mul*2*ref[w*ref_step]+add)>>shift),
-                 inverse);
-    }
-}
+            if(y){
+                if(prev_xc->x <= x)
+                    prev_xc++;
+                if(prev_xc->x == x + 1)
+                    rt= prev_xc->coeff;
+                else
+                    rt=0;
+            }
+            if(parent_xc){
+                if(x>>1 > parent_xc->x){
+                    parent_xc++;
+                }
+                if(x>>1 == parent_xc->x){
+                    p= parent_xc->coeff;
+                }
+            }
+            if(/*ll|*/l|lt|t|rt|p){
+                int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
 
-#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;
+                v=get_rac(&s->c, &b->state[0][context]);
+                if(v){
+                    v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
+                    v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + quant3bA[l&0xFF] + 3*quant3bA[t&0xFF]]);
 
-    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;
-    }
+                    xc->x=x;
+                    (xc++)->coeff= v;
+                }
+            }else{
+                if(!run){
+                    if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
+                    else           run= INT_MAX;
+                    v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
+                    v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
+
+                    xc->x=x;
+                    (xc++)->coeff= v;
+                }else{
+                    int max_run;
+                    run--;
+                    v=0;
+
+                    if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
+                    else  max_run= FFMIN(run, w-x-1);
+                    if(parent_xc)
+                        max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
+                    x+= max_run;
+                    run-= max_run;
+                }
+            }
+        }
+        (xc++)->x= w+1; //end marker
+        prev_xc= prev2_xc;
+        prev2_xc= xc;
 
-    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(parent_xc){
+            if(y&1){
+                while(parent_xc->x != parent->width+1)
+                    parent_xc++;
+                parent_xc++;
+                prev_parent_xc= parent_xc;
+            }else{
+                parent_xc= prev_parent_xc;
+            }
+        }
     }
 
-    if(mirror_right){
-        dst[w*dst_step] =
-            LIFTS(src[w*src_step], mul*2*ref[w*ref_step]+add, inverse);
-    }
+    (xc++)->x= w+1; //end marker
 }
-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;
 
-    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;
-    }
+static inline void decode_subband_slice_buffered(SnowContext *s, SubBand *b, slice_buffer * sb, int start_y, int h, int save_state[1]){
+    const int w= b->width;
+    int y;
+    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;
 
-    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(b->ibuf == s->spatial_idwt_buffer || s->qlog == LOSSLESS_QLOG){
+        qadd= 0;
+        qmul= 1<<QEXPSHIFT;
     }
 
-    if(mirror_right){
-        dst[w*dst_step] =
-            LIFTS(src[w*src_step], mul*2*ref[w*ref_step]+add, inverse);
-    }
-}
-#endif /* ! liftS */
+    /* If we are on the second or later slice, restore our index. */
+    if (start_y != 0)
+        new_index = save_state[0];
 
-static void horizontal_decompose53i(DWTELEM *b, int width){
-    DWTELEM temp[width];
-    const int width2= width>>1;
-    int x;
-    const int w2= (width+1)>>1;
 
-    for(x=0; x<width2; x++){
-        temp[x   ]= b[2*x    ];
-        temp[x+w2]= b[2*x + 1];
-    }
-    if(width&1)
-        temp[x   ]= b[2*x    ];
-#if 0
-    {
-    int A1,A2,A3,A4;
-    A2= temp[1       ];
-    A4= temp[0       ];
-    A1= temp[0+width2];
-    A1 -= (A2 + A4)>>1;
-    A4 += (A1 + 1)>>1;
-    b[0+width2] = A1;
-    b[0       ] = A4;
-    for(x=1; x+1<width2; x+=2){
-        A3= temp[x+width2];
-        A4= temp[x+1     ];
-        A3 -= (A2 + A4)>>1;
-        A2 += (A1 + A3 + 2)>>2;
-        b[x+width2] = A3;
-        b[x       ] = A2;
-
-        A1= temp[x+1+width2];
-        A2= temp[x+2       ];
-        A1 -= (A2 + A4)>>1;
-        A4 += (A1 + A3 + 2)>>2;
-        b[x+1+width2] = A1;
-        b[x+1       ] = A4;
-    }
-    A3= temp[width-1];
-    A3 -= A2;
-    A2 += (A1 + A3 + 2)>>2;
-    b[width -1] = A3;
-    b[width2-1] = A2;
+    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;
+            register int u= -(v&1);
+            line[x] = (t^u) - u;
+
+            v = b->x_coeff[new_index].coeff;
+            x = b->x_coeff[new_index++].x;
+        }
     }
-#else
-    lift(b+w2, temp+w2, temp, 1, 1, 1, width, -1, 0, 1, 1, 0);
-    lift(b   , temp   , b+w2, 1, 1, 1, width,  1, 2, 2, 0, 0);
-#endif /* 0 */
-}
 
-static void vertical_decompose53iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
-    int i;
+    /* Save our variables for the next slice. */
+    save_state[0] = new_index;
 
-    for(i=0; i<width; i++){
-        b1[i] -= (b0[i] + b2[i])>>1;
-    }
+    return;
 }
 
-static void vertical_decompose53iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
-    int i;
+static void reset_contexts(SnowContext *s){ //FIXME better initial contexts
+    int plane_index, level, orientation;
 
-    for(i=0; i<width; i++){
-        b1[i] += (b0[i] + b2[i] + 2)>>2;
+    for(plane_index=0; plane_index<3; plane_index++){
+        for(level=0; level<MAX_DECOMPOSITIONS; level++){
+            for(orientation=level ? 1:0; orientation<4; orientation++){
+                memset(s->plane[plane_index].band[level][orientation].state, MID_STATE, sizeof(s->plane[plane_index].band[level][orientation].state));
+            }
+        }
     }
+    memset(s->header_state, MID_STATE, sizeof(s->header_state));
+    memset(s->block_state, MID_STATE, sizeof(s->block_state));
 }
 
-static void spatial_decompose53i(DWTELEM *buffer, int width, int height, int stride){
-    int y;
-    DWTELEM *b0= buffer + mirror(-2-1, height-1)*stride;
-    DWTELEM *b1= buffer + mirror(-2  , height-1)*stride;
-
-    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);
+static int alloc_blocks(SnowContext *s){
+    int w= -((-s->avctx->width )>>LOG2_MB_SIZE);
+    int h= -((-s->avctx->height)>>LOG2_MB_SIZE);
 
-        if(y+1<(unsigned)height) vertical_decompose53iH0(b1, b2, b3, width);
-        if(y+0<(unsigned)height) vertical_decompose53iL0(b0, b1, b2, width);
+    s->b_width = w;
+    s->b_height= h;
 
-        b0=b2;
-        b1=b3;
-    }
+    av_free(s->block);
+    s->block= av_mallocz(w * h * sizeof(BlockNode) << (s->block_max_depth*2));
+    return 0;
 }
 
-static void horizontal_decompose97i(DWTELEM *b, int width){
-    DWTELEM temp[width];
-    const int w2= (width+1)>>1;
-
-    lift (temp+w2, b    +1, b      , 1, 2, 2, width,  W_AM, W_AO, W_AS, 1, 1);
-    liftS(temp   , b      , temp+w2, 1, 2, 1, width,  W_BM, W_BO, W_BS, 0, 0);
-    lift (b   +w2, temp+w2, temp   , 1, 1, 1, width,  W_CM, W_CO, W_CS, 1, 0);
-    lift (b      , temp   , b   +w2, 1, 1, 1, width,  W_DM, W_DO, W_DS, 0, 0);
+static inline void copy_rac_state(RangeCoder *d, RangeCoder *s){
+    uint8_t *bytestream= d->bytestream;
+    uint8_t *bytestream_start= d->bytestream_start;
+    *d= *s;
+    d->bytestream= bytestream;
+    d->bytestream_start= bytestream_start;
 }
 
+static inline void set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int ref, int type){
+    const int w= s->b_width << s->block_max_depth;
+    const int rem_depth= s->block_max_depth - level;
+    const int index= (x + y*w) << rem_depth;
+    const int block_w= 1<<rem_depth;
+    BlockNode block;
+    int i,j;
 
-static void vertical_decompose97iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
-    int i;
+    block.color[0]= l;
+    block.color[1]= cb;
+    block.color[2]= cr;
+    block.mx= mx;
+    block.my= my;
+    block.ref= ref;
+    block.type= type;
+    block.level= level;
 
-    for(i=0; i<width; i++){
-        b1[i] -= (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
+    for(j=0; j<block_w; j++){
+        for(i=0; i<block_w; i++){
+            s->block[index + i + j*w]= block;
+        }
     }
 }
 
-static void vertical_decompose97iH1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){
+static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
+    const int offset[3]= {
+          y*c->  stride + x,
+        ((y*c->uvstride + x)>>1),
+        ((y*c->uvstride + x)>>1),
+    };
     int i;
-
-    for(i=0; i<width; i++){
-        b1[i] += (W_CM*(b0[i] + b2[i])+W_CO)>>W_CS;
+    for(i=0; i<3; i++){
+        c->src[0][i]= src [i];
+        c->ref[0][i]= ref [i] + offset[i];
     }
+    assert(!ref_index);
 }
 
-static void vertical_decompose97iL0(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;
-#else
-        b1[i] = (16*4*b1[i] - 4*(b0[i] + b2[i]) + W_BO*5 + (5<<27)) / (5*16) - (1<<23);
-#endif
+static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
+                           const BlockNode *left, const BlockNode *top, const BlockNode *tr){
+    if(s->ref_frames == 1){
+        *mx = mid_pred(left->mx, top->mx, tr->mx);
+        *my = mid_pred(left->my, top->my, tr->my);
+    }else{
+        const int *scale = scale_mv_ref[ref];
+        *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
+                       (top ->mx * scale[top ->ref] + 128) >>8,
+                       (tr  ->mx * scale[tr  ->ref] + 128) >>8);
+        *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
+                       (top ->my * scale[top ->ref] + 128) >>8,
+                       (tr  ->my * scale[tr  ->ref] + 128) >>8);
     }
 }
 
-static void vertical_decompose97iL1(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;
+static av_always_inline int same_block(BlockNode *a, BlockNode *b){
+    if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
+        return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
+    }else{
+        return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
     }
 }
 
-static void spatial_decompose97i(DWTELEM *buffer, int width, int height, int stride){
-    int y;
-    DWTELEM *b0= buffer + mirror(-4-1, height-1)*stride;
-    DWTELEM *b1= buffer + mirror(-4  , height-1)*stride;
-    DWTELEM *b2= buffer + mirror(-4+1, height-1)*stride;
-    DWTELEM *b3= buffer + mirror(-4+2, height-1)*stride;
-
-    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;
+static void decode_q_branch(SnowContext *s, int level, int x, int y){
+    const int w= s->b_width << s->block_max_depth;
+    const int rem_depth= s->block_max_depth - level;
+    const int index= (x + y*w) << rem_depth;
+    int trx= (x+1)<<rem_depth;
+    const BlockNode *left  = x ? &s->block[index-1] : &null_block;
+    const BlockNode *top   = y ? &s->block[index-w] : &null_block;
+    const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
+    const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
+    int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
+
+    if(s->keyframe){
+        set_blocks(s, level, x, y, null_block.color[0], null_block.color[1], null_block.color[2], null_block.mx, null_block.my, null_block.ref, BLOCK_INTRA);
+        return;
     }
-}
 
-void ff_spatial_dwt(DWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count){
-    int level;
+    if(level==s->block_max_depth || get_rac(&s->c, &s->block_state[4 + s_context])){
+        int type, mx, my;
+        int l = left->color[0];
+        int cb= left->color[1];
+        int cr= left->color[2];
+        int ref = 0;
+        int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
+        int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 0*av_log2(2*FFABS(tr->mx - top->mx));
+        int my_context= av_log2(2*FFABS(left->my - top->my)) + 0*av_log2(2*FFABS(tr->my - top->my));
+
+        type= get_rac(&s->c, &s->block_state[1 + left->type + top->type]) ? BLOCK_INTRA : 0;
 
-    for(level=0; level<decomposition_count; level++){
-        switch(type){
-        case DWT_97: spatial_decompose97i(buffer, width>>level, height>>level, stride<<level); break;
-        case DWT_53: spatial_decompose53i(buffer, width>>level, height>>level, stride<<level); break;
+        if(type){
+            pred_mv(s, &mx, &my, 0, left, top, tr);
+            l += get_symbol(&s->c, &s->block_state[32], 1);
+            cb+= get_symbol(&s->c, &s->block_state[64], 1);
+            cr+= get_symbol(&s->c, &s->block_state[96], 1);
+        }else{
+            if(s->ref_frames > 1)
+                ref= get_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], 0);
+            pred_mv(s, &mx, &my, ref, left, top, tr);
+            mx+= get_symbol(&s->c, &s->block_state[128 + 32*(mx_context + 16*!!ref)], 1);
+            my+= get_symbol(&s->c, &s->block_state[128 + 32*(my_context + 16*!!ref)], 1);
         }
+        set_blocks(s, level, x, y, l, cb, cr, mx, my, ref, type);
+    }else{
+        decode_q_branch(s, level+1, 2*x+0, 2*y+0);
+        decode_q_branch(s, level+1, 2*x+1, 2*y+0);
+        decode_q_branch(s, level+1, 2*x+0, 2*y+1);
+        decode_q_branch(s, level+1, 2*x+1, 2*y+1);
     }
 }
 
-static void horizontal_compose53i(IDWTELEM *b, int width){
-    IDWTELEM temp[width];
-    const int width2= width>>1;
-    const int w2= (width+1)>>1;
-    int x;
+static void decode_blocks(SnowContext *s){
+    int x, y;
+    int w= s->b_width;
+    int h= s->b_height;
 
-#if 0
-    int A1,A2,A3,A4;
-    A2= temp[1       ];
-    A4= temp[0       ];
-    A1= temp[0+width2];
-    A1 -= (A2 + A4)>>1;
-    A4 += (A1 + 1)>>1;
-    b[0+width2] = A1;
-    b[0       ] = A4;
-    for(x=1; x+1<width2; x+=2){
-        A3= temp[x+width2];
-        A4= temp[x+1     ];
-        A3 -= (A2 + A4)>>1;
-        A2 += (A1 + A3 + 2)>>2;
-        b[x+width2] = A3;
-        b[x       ] = A2;
-
-        A1= temp[x+1+width2];
-        A2= temp[x+2       ];
-        A1 -= (A2 + A4)>>1;
-        A4 += (A1 + A3 + 2)>>2;
-        b[x+1+width2] = A1;
-        b[x+1       ] = A4;
-    }
-    A3= temp[width-1];
-    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 /* 0 */
-    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;
-    }
-}
-
-static void spatial_compose53i_buffered_init(DWTCompose *cs, slice_buffer * sb, int height, int stride_line){
-    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(DWTCompose *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(DWTCompose *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);
-
-        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;
-}
-
-static void spatial_compose53i_dy(DWTCompose *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;
-
-        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;
-}
-
-static void av_unused spatial_compose53i(IDWTELEM *buffer, int width, int height, int stride){
-    DWTCompose 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_lift (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++){
-        b1[i] -= (W_CM*(b0[i] + b2[i])+W_CO)>>W_CS;
-    }
-}
-
-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;
-#else
-        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;
+    for(y=0; y<h; y++){
+        for(x=0; x<w; x++){
+            decode_q_branch(s, 0, x, y);
+        }
     }
 }
 
-void ff_snow_vertical_compose97i(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
-    int i;
+static void mc_block(Plane *p, uint8_t *dst, const uint8_t *src, int stride, int b_w, int b_h, int dx, int dy){
+    static const uint8_t weight[64]={
+    8,7,6,5,4,3,2,1,
+    7,7,0,0,0,0,0,1,
+    6,0,6,0,0,0,2,0,
+    5,0,0,5,0,3,0,0,
+    4,0,0,0,4,0,0,0,
+    3,0,0,5,0,3,0,0,
+    2,0,6,0,0,0,2,0,
+    1,7,0,0,0,0,0,1,
+    };
 
-    for(i=0; i<width; i++){
-        b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
-        b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
-#ifdef liftS
-        b2[i] += (W_BM*(b1[i] + b3[i])+W_BO)>>W_BS;
-#else
-        b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
-#endif
-        b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
-    }
-}
+    static const uint8_t brane[256]={
+    0x00,0x01,0x01,0x01,0x01,0x01,0x01,0x01,0x11,0x12,0x12,0x12,0x12,0x12,0x12,0x12,
+    0x04,0x05,0xcc,0xcc,0xcc,0xcc,0xcc,0x41,0x15,0x16,0xcc,0xcc,0xcc,0xcc,0xcc,0x52,
+    0x04,0xcc,0x05,0xcc,0xcc,0xcc,0x41,0xcc,0x15,0xcc,0x16,0xcc,0xcc,0xcc,0x52,0xcc,
+    0x04,0xcc,0xcc,0x05,0xcc,0x41,0xcc,0xcc,0x15,0xcc,0xcc,0x16,0xcc,0x52,0xcc,0xcc,
+    0x04,0xcc,0xcc,0xcc,0x41,0xcc,0xcc,0xcc,0x15,0xcc,0xcc,0xcc,0x16,0xcc,0xcc,0xcc,
+    0x04,0xcc,0xcc,0x41,0xcc,0x05,0xcc,0xcc,0x15,0xcc,0xcc,0x52,0xcc,0x16,0xcc,0xcc,
+    0x04,0xcc,0x41,0xcc,0xcc,0xcc,0x05,0xcc,0x15,0xcc,0x52,0xcc,0xcc,0xcc,0x16,0xcc,
+    0x04,0x41,0xcc,0xcc,0xcc,0xcc,0xcc,0x05,0x15,0x52,0xcc,0xcc,0xcc,0xcc,0xcc,0x16,
+    0x44,0x45,0x45,0x45,0x45,0x45,0x45,0x45,0x55,0x56,0x56,0x56,0x56,0x56,0x56,0x56,
+    0x48,0x49,0xcc,0xcc,0xcc,0xcc,0xcc,0x85,0x59,0x5A,0xcc,0xcc,0xcc,0xcc,0xcc,0x96,
+    0x48,0xcc,0x49,0xcc,0xcc,0xcc,0x85,0xcc,0x59,0xcc,0x5A,0xcc,0xcc,0xcc,0x96,0xcc,
+    0x48,0xcc,0xcc,0x49,0xcc,0x85,0xcc,0xcc,0x59,0xcc,0xcc,0x5A,0xcc,0x96,0xcc,0xcc,
+    0x48,0xcc,0xcc,0xcc,0x49,0xcc,0xcc,0xcc,0x59,0xcc,0xcc,0xcc,0x96,0xcc,0xcc,0xcc,
+    0x48,0xcc,0xcc,0x85,0xcc,0x49,0xcc,0xcc,0x59,0xcc,0xcc,0x96,0xcc,0x5A,0xcc,0xcc,
+    0x48,0xcc,0x85,0xcc,0xcc,0xcc,0x49,0xcc,0x59,0xcc,0x96,0xcc,0xcc,0xcc,0x5A,0xcc,
+    0x48,0x85,0xcc,0xcc,0xcc,0xcc,0xcc,0x49,0x59,0x96,0xcc,0xcc,0xcc,0xcc,0xcc,0x5A,
+    };
 
-static void spatial_compose97i_buffered_init(DWTCompose *cs, slice_buffer * sb, int height, int stride_line){
-    cs->b0 = slice_buffer_get_line(sb, mirror(-3-1, height-1) * stride_line);
-    cs->b1 = slice_buffer_get_line(sb, mirror(-3  , height-1) * stride_line);
-    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 const uint8_t needs[16]={
+    0,1,0,0,
+    2,4,2,0,
+    0,1,0,0,
+    15
+    };
 
-static void spatial_compose97i_init(DWTCompose *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;
-}
+    int x, y, b, r, l;
+    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;
 
-static void spatial_compose97i_dy_buffered(DSPContext *dsp, DWTCompose *cs, slice_buffer * sb, int width, int height, int stride_line){
-    int y = cs->y;
+    b= needs[l] | needs[r];
+    if(p && !p->diag_mc)
+        b= 15;
 
-    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);
+    if(b&5){
+        for(y=0; y < b_h+HTAPS_MAX-1; y++){
+            for(x=0; x < b_w; x++){
+                int a_1=src[x + HTAPS_MAX/2-4];
+                int a0= src[x + HTAPS_MAX/2-3];
+                int a1= src[x + HTAPS_MAX/2-2];
+                int a2= src[x + HTAPS_MAX/2-1];
+                int a3= src[x + HTAPS_MAX/2+0];
+                int a4= src[x + HTAPS_MAX/2+1];
+                int a5= src[x + HTAPS_MAX/2+2];
+                int a6= src[x + HTAPS_MAX/2+3];
+                int am=0;
+                if(!p || p->fast_mc){
+                    am= 20*(a2+a3) - 5*(a1+a4) + (a0+a5);
+                    tmpI[x]= am;
+                    am= (am+16)>>5;
+                }else{
+                    am= p->hcoeff[0]*(a2+a3) + p->hcoeff[1]*(a1+a4) + p->hcoeff[2]*(a0+a5) + p->hcoeff[3]*(a_1+a6);
+                    tmpI[x]= am;
+                    am= (am+32)>>6;
+                }
 
-    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(am&(~255)) am= ~(am>>31);
+                tmp2[x]= am;
+            }
+            tmpI+= 64;
+            tmp2+= stride;
+            src += stride;
+        }
+        src -= stride*y;
     }
+    src += HTAPS_MAX/2 - 1;
+    tmp2= tmp2t[1];
 
-        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;
-    cs->y += 2;
-}
-
-static void spatial_compose97i_dy(DWTCompose *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;
-
-        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;
-    cs->y += 2;
-}
-
-static void av_unused spatial_compose97i(IDWTELEM *buffer, int width, int height, int stride){
-    DWTCompose cs;
-    spatial_compose97i_init(&cs, buffer, height, stride);
-    while(cs.y <= height)
-        spatial_compose97i_dy(&cs, buffer, width, height, stride);
-}
+    if(b&2){
+        for(y=0; y < b_h; y++){
+            for(x=0; x < b_w+1; x++){
+                int a_1=src[x + (HTAPS_MAX/2-4)*stride];
+                int a0= src[x + (HTAPS_MAX/2-3)*stride];
+                int a1= src[x + (HTAPS_MAX/2-2)*stride];
+                int a2= src[x + (HTAPS_MAX/2-1)*stride];
+                int a3= src[x + (HTAPS_MAX/2+0)*stride];
+                int a4= src[x + (HTAPS_MAX/2+1)*stride];
+                int a5= src[x + (HTAPS_MAX/2+2)*stride];
+                int a6= src[x + (HTAPS_MAX/2+3)*stride];
+                int am=0;
+                if(!p || p->fast_mc)
+                    am= (20*(a2+a3) - 5*(a1+a4) + (a0+a5) + 16)>>5;
+                else
+                    am= (p->hcoeff[0]*(a2+a3) + p->hcoeff[1]*(a1+a4) + p->hcoeff[2]*(a0+a5) + p->hcoeff[3]*(a_1+a6) + 32)>>6;
 
-static void ff_spatial_idwt_buffered_init(DWTCompose *cs, slice_buffer * sb, int width, int height, int stride_line, int type, int decomposition_count){
-    int level;
-    for(level=decomposition_count-1; level>=0; level--){
-        switch(type){
-        case DWT_97: spatial_compose97i_buffered_init(cs+level, sb, height>>level, stride_line<<level); break;
-        case DWT_53: spatial_compose53i_buffered_init(cs+level, sb, height>>level, stride_line<<level); break;
+                if(am&(~255)) am= ~(am>>31);
+                tmp2[x]= am;
+            }
+            src += stride;
+            tmp2+= stride;
         }
+        src -= stride*y;
     }
-}
-
-static void ff_spatial_idwt_init(DWTCompose *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;
+    src += stride*(HTAPS_MAX/2 - 1);
+    tmp2= tmp2t[2];
+    tmpI= tmpIt;
+    if(b&4){
+        for(y=0; y < b_h; y++){
+            for(x=0; x < b_w; x++){
+                int a_1=tmpI[x + (HTAPS_MAX/2-4)*64];
+                int a0= tmpI[x + (HTAPS_MAX/2-3)*64];
+                int a1= tmpI[x + (HTAPS_MAX/2-2)*64];
+                int a2= tmpI[x + (HTAPS_MAX/2-1)*64];
+                int a3= tmpI[x + (HTAPS_MAX/2+0)*64];
+                int a4= tmpI[x + (HTAPS_MAX/2+1)*64];
+                int a5= tmpI[x + (HTAPS_MAX/2+2)*64];
+                int a6= tmpI[x + (HTAPS_MAX/2+3)*64];
+                int am=0;
+                if(!p || p->fast_mc)
+                    am= (20*(a2+a3) - 5*(a1+a4) + (a0+a5) + 512)>>10;
+                else
+                    am= (p->hcoeff[0]*(a2+a3) + p->hcoeff[1]*(a1+a4) + p->hcoeff[2]*(a0+a5) + p->hcoeff[3]*(a_1+a6) + 2048)>>12;
+                if(am&(~255)) am= ~(am>>31);
+                tmp2[x]= am;
+            }
+            tmpI+= 64;
+            tmp2+= stride;
         }
     }
-}
 
-static void ff_spatial_idwt_slice(DWTCompose *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;
+    hpel[ 0]= src;
+    hpel[ 1]= tmp2t[0] + stride*(HTAPS_MAX/2-1);
+    hpel[ 2]= src + 1;
 
-    for(level=decomposition_count-1; level>=0; level--){
-        while(cs[level].y <= FFMIN((y>>level)+support, height>>level)){
-            switch(type){
-            case DWT_97: spatial_compose97i_dy(cs+level, buffer, width>>level, height>>level, stride<<level);
-                break;
-            case DWT_53: spatial_compose53i_dy(cs+level, buffer, width>>level, height>>level, stride<<level);
-                break;
-            }
-        }
-    }
-}
+    hpel[ 4]= tmp2t[1];
+    hpel[ 5]= tmp2t[2];
+    hpel[ 6]= tmp2t[1] + 1;
 
-static void ff_spatial_idwt_buffered_slice(DSPContext *dsp, DWTCompose *cs, slice_buffer * slice_buf, int width, int height, int stride_line, int type, int decomposition_count, int y){
-    const int support = type==1 ? 3 : 5;
-    int level;
-    if(type==2) return;
+    hpel[ 8]= src + stride;
+    hpel[ 9]= hpel[1] + stride;
+    hpel[10]= hpel[8] + 1;
 
-    for(level=decomposition_count-1; level>=0; level--){
-        while(cs[level].y <= FFMIN((y>>level)+support, height>>level)){
-            switch(type){
-            case DWT_97: spatial_compose97i_dy_buffered(dsp, cs+level, slice_buf, width>>level, height>>level, stride_line<<level);
-                break;
-            case DWT_53: spatial_compose53i_dy_buffered(cs+level, slice_buf, width>>level, height>>level, stride_line<<level);
-                break;
+    if(b==15){
+        const uint8_t *src1= hpel[dx/8 + dy/8*4  ];
+        const uint8_t *src2= hpel[dx/8 + dy/8*4+1];
+        const uint8_t *src3= hpel[dx/8 + dy/8*4+4];
+        const uint8_t *src4= hpel[dx/8 + dy/8*4+5];
+        dx&=7;
+        dy&=7;
+        for(y=0; y < b_h; y++){
+            for(x=0; x < b_w; x++){
+                dst[x]= ((8-dx)*(8-dy)*src1[x] + dx*(8-dy)*src2[x]+
+                         (8-dx)*   dy *src3[x] + dx*   dy *src4[x]+32)>>6;
+            }
+            src1+=stride;
+            src2+=stride;
+            src3+=stride;
+            src4+=stride;
+            dst +=stride;
+        }
+    }else{
+        const uint8_t *src1= hpel[l];
+        const uint8_t *src2= hpel[r];
+        int a= weight[((dx&7) + (8*(dy&7)))];
+        int b= 8-a;
+        for(y=0; y < b_h; y++){
+            for(x=0; x < b_w; x++){
+                dst[x]= (a*src1[x] + b*src2[x] + 4)>>3;
             }
+            src1+=stride;
+            src2+=stride;
+            dst +=stride;
         }
     }
 }
 
-static void ff_spatial_idwt(IDWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count){
-        DWTCompose cs[MAX_DECOMPOSITIONS];
-        int y;
-        ff_spatial_idwt_init(cs, buffer, width, height, stride, type, decomposition_count);
-        for(y=0; y<height; y+=4)
-            ff_spatial_idwt_slice(cs, buffer, width, height, stride, type, decomposition_count, y);
+#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){\
+    assert(h==b_w);\
+    mc_block(NULL, dst, src-(HTAPS_MAX/2-1)-(HTAPS_MAX/2-1)*stride, stride, b_w, b_w, dx, dy);\
 }
 
-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){
-        int run=0;
-        int runs[w*h];
-        int run_index=0;
-        int max_index;
-
-        for(y=0; y<h; y++){
-            for(x=0; x<w; x++){
-                int v, p=0;
-                int /*ll=0, */l=0, lt=0, t=0, rt=0;
-                v= src[x + y*stride];
+mca( 0, 0,16)
+mca( 8, 0,16)
+mca( 0, 8,16)
+mca( 8, 8,16)
+mca( 0, 0,8)
+mca( 8, 0,8)
+mca( 0, 8,8)
+mca( 8, 8,8)
 
-                if(y){
-                    t= src[x + (y-1)*stride];
-                    if(x){
-                        lt= src[x - 1 + (y-1)*stride];
-                    }
-                    if(x + 1 < w){
-                        rt= src[x + 1 + (y-1)*stride];
-                    }
-                }
-                if(x){
-                    l= src[x - 1 + y*stride];
-                    /*if(x > 1){
-                        if(orientation==1) ll= src[y + (x-2)*stride];
-                        else               ll= src[x - 2 + y*stride];
-                    }*/
-                }
-                if(parent){
-                    int px= x>>1;
-                    int py= y>>1;
-                    if(px<b->parent->width && py<b->parent->height)
-                        p= parent[px + py*2*stride];
-                }
-                if(!(/*ll|*/l|lt|t|rt|p)){
-                    if(v){
-                        runs[run_index++]= run;
-                        run=0;
-                    }else{
-                        run++;
-                    }
+static void pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, int stride, int sx, int sy, int b_w, int b_h, BlockNode *block, int plane_index, int w, int h){
+    if(block->type & BLOCK_INTRA){
+        int x, y;
+        const int color = block->color[plane_index];
+        const int color4= color*0x01010101;
+        if(b_w==32){
+            for(y=0; y < b_h; y++){
+                *(uint32_t*)&dst[0 + y*stride]= color4;
+                *(uint32_t*)&dst[4 + y*stride]= color4;
+                *(uint32_t*)&dst[8 + y*stride]= color4;
+                *(uint32_t*)&dst[12+ y*stride]= color4;
+                *(uint32_t*)&dst[16+ y*stride]= color4;
+                *(uint32_t*)&dst[20+ y*stride]= color4;
+                *(uint32_t*)&dst[24+ y*stride]= color4;
+                *(uint32_t*)&dst[28+ y*stride]= color4;
+            }
+        }else if(b_w==16){
+            for(y=0; y < b_h; y++){
+                *(uint32_t*)&dst[0 + y*stride]= color4;
+                *(uint32_t*)&dst[4 + y*stride]= color4;
+                *(uint32_t*)&dst[8 + y*stride]= color4;
+                *(uint32_t*)&dst[12+ y*stride]= color4;
+            }
+        }else if(b_w==8){
+            for(y=0; y < b_h; y++){
+                *(uint32_t*)&dst[0 + y*stride]= color4;
+                *(uint32_t*)&dst[4 + y*stride]= color4;
+            }
+        }else if(b_w==4){
+            for(y=0; y < b_h; y++){
+                *(uint32_t*)&dst[0 + y*stride]= color4;
+            }
+        }else{
+            for(y=0; y < b_h; y++){
+                for(x=0; x < b_w; x++){
+                    dst[x + y*stride]= color;
                 }
             }
         }
-        max_index= run_index;
-        runs[run_index++]= run;
-        run_index=0;
-        run= runs[run_index++];
-
-        put_symbol2(&s->c, b->state[30], max_index, 0);
-        if(run_index <= max_index)
-            put_symbol2(&s->c, b->state[1], run, 3);
-
-        for(y=0; y<h; y++){
-            if(s->c.bytestream_end - s->c.bytestream < w*40){
-                av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
-                return -1;
+    }else{
+        uint8_t *src= s->last_picture[block->ref].data[plane_index];
+        const int scale= plane_index ?  s->mv_scale : 2*s->mv_scale;
+        int mx= block->mx*scale;
+        int my= block->my*scale;
+        const int dx= mx&15;
+        const int dy= my&15;
+        const int tab_index= 3 - (b_w>>2) + (b_w>>4);
+        sx += (mx>>4) - (HTAPS_MAX/2-1);
+        sy += (my>>4) - (HTAPS_MAX/2-1);
+        src += sx + sy*stride;
+        if(   (unsigned)sx >= w - b_w - (HTAPS_MAX-2)
+           || (unsigned)sy >= h - b_h - (HTAPS_MAX-2)){
+            s->dsp.emulated_edge_mc(tmp + MB_SIZE, src, stride, b_w+HTAPS_MAX-1, b_h+HTAPS_MAX-1, sx, sy, w, h);
+            src= tmp + MB_SIZE;
+        }
+//        assert(b_w == b_h || 2*b_w == b_h || b_w == 2*b_h);
+//        assert(!(b_w&(b_w-1)));
+        assert(b_w>1 && b_h>1);
+        assert((tab_index>=0 && tab_index<4) || b_w==32);
+        if((dx&3) || (dy&3) || !(b_w == b_h || 2*b_w == b_h || b_w == 2*b_h) || (b_w&(b_w-1)) || !s->plane[plane_index].fast_mc )
+            mc_block(&s->plane[plane_index], dst, src, stride, b_w, b_h, dx, dy);
+        else if(b_w==32){
+            int y;
+            for(y=0; y<b_h; y+=16){
+                s->dsp.put_h264_qpel_pixels_tab[0][dy+(dx>>2)](dst + y*stride, src + 3 + (y+3)*stride,stride);
+                s->dsp.put_h264_qpel_pixels_tab[0][dy+(dx>>2)](dst + 16 + y*stride, src + 19 + (y+3)*stride,stride);
             }
-            for(x=0; x<w; x++){
-                int v, p=0;
-                int /*ll=0, */l=0, lt=0, t=0, rt=0;
-                v= src[x + y*stride];
-
-                if(y){
-                    t= src[x + (y-1)*stride];
-                    if(x){
-                        lt= src[x - 1 + (y-1)*stride];
-                    }
-                    if(x + 1 < w){
-                        rt= src[x + 1 + (y-1)*stride];
-                    }
-                }
-                if(x){
-                    l= src[x - 1 + y*stride];
-                    /*if(x > 1){
-                        if(orientation==1) ll= src[y + (x-2)*stride];
-                        else               ll= src[x - 2 + y*stride];
-                    }*/
-                }
-                if(parent){
-                    int px= x>>1;
-                    int py= y>>1;
-                    if(px<b->parent->width && py<b->parent->height)
-                        p= parent[px + py*2*stride];
-                }
-                if(/*ll|*/l|lt|t|rt|p){
-                    int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
-
-                    put_rac(&s->c, &b->state[0][context], !!v);
-                }else{
-                    if(!run){
-                        run= runs[run_index++];
+        }else if(b_w==b_h)
+            s->dsp.put_h264_qpel_pixels_tab[tab_index  ][dy+(dx>>2)](dst,src + 3 + 3*stride,stride);
+        else if(b_w==2*b_h){
+            s->dsp.put_h264_qpel_pixels_tab[tab_index+1][dy+(dx>>2)](dst    ,src + 3       + 3*stride,stride);
+            s->dsp.put_h264_qpel_pixels_tab[tab_index+1][dy+(dx>>2)](dst+b_h,src + 3 + b_h + 3*stride,stride);
+        }else{
+            assert(2*b_w==b_h);
+            s->dsp.put_h264_qpel_pixels_tab[tab_index  ][dy+(dx>>2)](dst           ,src + 3 + 3*stride           ,stride);
+            s->dsp.put_h264_qpel_pixels_tab[tab_index  ][dy+(dx>>2)](dst+b_w*stride,src + 3 + 3*stride+b_w*stride,stride);
+        }
+    }
+}
 
-                        if(run_index <= max_index)
-                            put_symbol2(&s->c, b->state[1], run, 3);
-                        assert(v);
-                    }else{
-                        run--;
-                        assert(!v);
-                    }
-                }
-                if(v){
-                    int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
-                    int l2= 2*FFABS(l) + (l<0);
-                    int t2= 2*FFABS(t) + (t<0);
+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);
+        const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
+        dst = slice_buffer_get_line(sb, src_y + y);
+        for(x=0; x<b_w; x++){
+            int v=   obmc1[x] * block[3][x + y*src_stride]
+                    +obmc2[x] * block[2][x + y*src_stride]
+                    +obmc3[x] * block[1][x + y*src_stride]
+                    +obmc4[x] * block[0][x + y*src_stride];
 
-                    put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
-                    put_rac(&s->c, &b->state[0][16 + 1 + 3 + quant3bA[l2&0xFF] + 3*quant3bA[t2&0xFF]], v<0);
-                }
+            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;
+                if(v&(~255)) v= ~(v>>31);
+                dst8[x + y*src_stride] = v;
+            }else{
+                dst[x + src_x] -= v;
             }
         }
     }
-    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);
-}
-
-static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
-    const int w= b->width;
-    const int h= b->height;
+//FIXME name cleanup (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;
+    BlockNode *lb= lt+b_stride;
+    BlockNode *rb= lb+1;
+    uint8_t *block[4];
+    int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
+    uint8_t *tmp = s->scratchbuf;
+    uint8_t *ptmp;
     int x,y;
 
-    if(1){
-        int run, runs;
-        x_and_coeff *xc= b->x_coeff;
-        x_and_coeff *prev_xc= NULL;
-        x_and_coeff *prev2_xc= xc;
-        x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
-        x_and_coeff *prev_parent_xc= parent_xc;
-
-        runs= get_symbol2(&s->c, b->state[30], 0);
-        if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
-        else           run= INT_MAX;
+    if(b_x<0){
+        lt= rt;
+        lb= rb;
+    }else if(b_x + 1 >= b_width){
+        rt= lt;
+        rb= lb;
+    }
+    if(b_y<0){
+        lt= lb;
+        rt= rb;
+    }else if(b_y + 1 >= b_height){
+        lb= lt;
+        rb= rt;
+    }
 
-        for(y=0; y<h; y++){
-            int v=0;
-            int lt=0, t=0, rt=0;
+    if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
+        obmc -= src_x;
+        b_w += src_x;
+        if(!sliced && !offset_dst)
+            dst -= src_x;
+        src_x=0;
+    }else if(src_x + b_w > w){
+        b_w = w - src_x;
+    }
+    if(src_y<0){
+        obmc -= src_y*obmc_stride;
+        b_h += src_y;
+        if(!sliced && !offset_dst)
+            dst -= src_y*dst_stride;
+        src_y=0;
+    }else if(src_y + b_h> h){
+        b_h = h - src_y;
+    }
 
-            if(y && prev_xc->x == 0){
-                rt= prev_xc->coeff;
-            }
-            for(x=0; x<w; x++){
-                int p=0;
-                const int l= v;
+    if(b_w<=0 || b_h<=0) return;
 
-                lt= t; t= rt;
+    assert(src_stride > 2*MB_SIZE + 5);
 
-                if(y){
-                    if(prev_xc->x <= x)
-                        prev_xc++;
-                    if(prev_xc->x == x + 1)
-                        rt= prev_xc->coeff;
-                    else
-                        rt=0;
-                }
-                if(parent_xc){
-                    if(x>>1 > parent_xc->x){
-                        parent_xc++;
-                    }
-                    if(x>>1 == parent_xc->x){
-                        p= parent_xc->coeff;
-                    }
-                }
-                if(/*ll|*/l|lt|t|rt|p){
-                    int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
+    if(!sliced && offset_dst)
+        dst += src_x + src_y*dst_stride;
+    dst8+= src_x + src_y*src_stride;
+//    src += src_x + src_y*src_stride;
 
-                    v=get_rac(&s->c, &b->state[0][context]);
-                    if(v){
-                        v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
-                        v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + quant3bA[l&0xFF] + 3*quant3bA[t&0xFF]]);
+    ptmp= tmp + 3*tmp_step;
+    block[0]= ptmp;
+    ptmp+=tmp_step;
+    pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
 
-                        xc->x=x;
-                        (xc++)->coeff= v;
-                    }
-                }else{
-                    if(!run){
-                        if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
-                        else           run= INT_MAX;
-                        v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
-                        v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
+    if(same_block(lt, rt)){
+        block[1]= block[0];
+    }else{
+        block[1]= ptmp;
+        ptmp+=tmp_step;
+        pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
+    }
 
-                        xc->x=x;
-                        (xc++)->coeff= v;
-                    }else{
-                        int max_run;
-                        run--;
-                        v=0;
-
-                        if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
-                        else  max_run= FFMIN(run, w-x-1);
-                        if(parent_xc)
-                            max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
-                        x+= max_run;
-                        run-= max_run;
-                    }
-                }
-            }
-            (xc++)->x= w+1; //end marker
-            prev_xc= prev2_xc;
-            prev2_xc= xc;
+    if(same_block(lt, lb)){
+        block[2]= block[0];
+    }else if(same_block(rt, lb)){
+        block[2]= block[1];
+    }else{
+        block[2]= ptmp;
+        ptmp+=tmp_step;
+        pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
+    }
 
-            if(parent_xc){
-                if(y&1){
-                    while(parent_xc->x != parent->width+1)
-                        parent_xc++;
-                    parent_xc++;
-                    prev_parent_xc= parent_xc;
+    if(same_block(lt, rb) ){
+        block[3]= block[0];
+    }else if(same_block(rt, rb)){
+        block[3]= block[1];
+    }else if(same_block(lb, rb)){
+        block[3]= block[2];
+    }else{
+        block[3]= ptmp;
+        pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
+    }
+#if 0
+    for(y=0; y<b_h; y++){
+        for(x=0; x<b_w; x++){
+            int v=   obmc [x + y*obmc_stride] * block[3][x + y*src_stride] * (256/OBMC_MAX);
+            if(add) dst[x + y*dst_stride] += v;
+            else    dst[x + y*dst_stride] -= v;
+        }
+    }
+    for(y=0; y<b_h; y++){
+        uint8_t *obmc2= obmc + (obmc_stride>>1);
+        for(x=0; x<b_w; x++){
+            int v=   obmc2[x + y*obmc_stride] * block[2][x + y*src_stride] * (256/OBMC_MAX);
+            if(add) dst[x + y*dst_stride] += v;
+            else    dst[x + y*dst_stride] -= v;
+        }
+    }
+    for(y=0; y<b_h; y++){
+        uint8_t *obmc3= obmc + obmc_stride*(obmc_stride>>1);
+        for(x=0; x<b_w; x++){
+            int v=   obmc3[x + y*obmc_stride] * block[1][x + y*src_stride] * (256/OBMC_MAX);
+            if(add) dst[x + y*dst_stride] += v;
+            else    dst[x + y*dst_stride] -= v;
+        }
+    }
+    for(y=0; y<b_h; y++){
+        uint8_t *obmc3= obmc + obmc_stride*(obmc_stride>>1);
+        uint8_t *obmc4= obmc3+ (obmc_stride>>1);
+        for(x=0; x<b_w; x++){
+            int v=   obmc4[x + y*obmc_stride] * block[0][x + y*src_stride] * (256/OBMC_MAX);
+            if(add) dst[x + y*dst_stride] += v;
+            else    dst[x + y*dst_stride] -= v;
+        }
+    }
+#else
+    if(sliced){
+        s->dwt.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);
+            const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
+            const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
+            for(x=0; x<b_w; x++){
+                int v=   obmc1[x] * block[3][x + y*src_stride]
+                        +obmc2[x] * block[2][x + y*src_stride]
+                        +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;
+                    if(v&(~255)) v= ~(v>>31);
+                    dst8[x + y*src_stride] = v;
                 }else{
-                    parent_xc= prev_parent_xc;
+                    dst[x + y*dst_stride] -= v;
                 }
             }
         }
-
-        (xc++)->x= w+1; //end marker
     }
+#endif /* 0 */
 }
 
-static inline void decode_subband_slice_buffered(SnowContext *s, SubBand *b, slice_buffer * sb, int start_y, int h, int save_state[1]){
-    const int w= b->width;
-    int y;
-    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;
-    }
-
-    /* If we are on the second or later slice, restore our index. */
-    if (start_y != 0)
-        new_index = save_state[0];
-
+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;
+    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];
+    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;
 
-    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;
-            register int u= -(v&1);
-            line[x] = (t^u) - u;
+    if(s->keyframe || (s->avctx->debug&512)){
+        if(mb_y==mb_h)
+            return;
 
-            v = b->x_coeff[new_index].coeff;
-            x = b->x_coeff[new_index++].x;
+        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;
+                    if(v&(~255)) v= ~(v>>31);
+                    dst8[x + y*ref_stride]= v;
+                }
+            }
+        }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;
+                }
+            }
         }
-    }
 
-    /* Save our variables for the next slice. */
-    save_state[0] = new_index;
+        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 void reset_contexts(SnowContext *s){ //FIXME better initial contexts
-    int plane_index, level, orientation;
+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;
+    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;
+    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;
 
-    for(plane_index=0; plane_index<3; plane_index++){
-        for(level=0; level<MAX_DECOMPOSITIONS; level++){
-            for(orientation=level ? 1:0; orientation<4; orientation++){
-                memset(s->plane[plane_index].band[level][orientation].state, MID_STATE, sizeof(s->plane[plane_index].band[level][orientation].state));
+    if(s->keyframe || (s->avctx->debug&512)){
+        if(mb_y==mb_h)
+            return;
+
+        if(add){
+            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
+                for(x=0; x<w; x++){
+                    int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
+                    v >>= FRAC_BITS;
+                    if(v&(~255)) v= ~(v>>31);
+                    dst8[x + y*ref_stride]= v;
+                }
+            }
+        }else{
+            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
+                for(x=0; x<w; x++){
+                    buf[x + y*w]-= 128<<FRAC_BITS;
+                }
             }
         }
+
+        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);
     }
-    memset(s->header_state, MID_STATE, sizeof(s->header_state));
-    memset(s->block_state, MID_STATE, sizeof(s->block_state));
 }
 
-static int alloc_blocks(SnowContext *s){
-    int w= -((-s->avctx->width )>>LOG2_MB_SIZE);
-    int h= -((-s->avctx->height)>>LOG2_MB_SIZE);
+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);
+}
 
-    s->b_width = w;
-    s->b_height= h;
+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;
 
-    s->block= av_mallocz(w * h * sizeof(BlockNode) << (s->block_max_depth*2));
-    return 0;
-}
+    if(s->qlog == LOSSLESS_QLOG) return;
 
-static inline void copy_rac_state(RangeCoder *d, RangeCoder *s){
-    uint8_t *bytestream= d->bytestream;
-    uint8_t *bytestream_start= d->bytestream_start;
-    *d= *s;
-    d->bytestream= bytestream;
-    d->bytestream_start= bytestream_start;
+    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){
+                line[x]=  (( i*qmul + qadd)>>(QEXPSHIFT));
+            }
+        }
+    }
 }
 
-//near copy & paste from dsputil, FIXME
-static int pix_sum(uint8_t * pix, int line_size, int w)
-{
-    int s, i, j;
+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;
 
-    s = 0;
-    for (i = 0; i < w; i++) {
-        for (j = 0; j < w; j++) {
-            s += pix[0];
-            pix ++;
+    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++){
+        prev = line;
+//        line = slice_buffer_get_line_from_address(sb, src + (y * stride));
+        line = slice_buffer_get_line(sb, (y * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;
+        for(x=0; x<w; x++){
+            if(x){
+                if(use_median){
+                    if(y && x+1<w) line[x] += mid_pred(line[x - 1], prev[x], prev[x + 1]);
+                    else  line[x] += line[x - 1];
+                }else{
+                    if(y) line[x] += mid_pred(line[x - 1], prev[x], line[x - 1] + prev[x] - prev[x - 1]);
+                    else  line[x] += line[x - 1];
+                }
+            }else{
+                if(y) line[x] += prev[x];
+            }
         }
-        pix += line_size - w;
     }
-    return s;
 }
 
-//near copy & paste from dsputil, FIXME
-static int pix_norm1(uint8_t * pix, int line_size, int w)
-{
-    int s, i, j;
-    uint32_t *sq = ff_squareTbl + 256;
+static void decode_qlogs(SnowContext *s){
+    int plane_index, level, orientation;
 
-    s = 0;
-    for (i = 0; i < w; i++) {
-        for (j = 0; j < w; j ++) {
-            s += sq[pix[0]];
-            pix ++;
+    for(plane_index=0; plane_index<3; plane_index++){
+        for(level=0; level<s->spatial_decomposition_count; level++){
+            for(orientation=level ? 1:0; orientation<4; orientation++){
+                int q;
+                if     (plane_index==2) q= s->plane[1].band[level][orientation].qlog;
+                else if(orientation==2) q= s->plane[plane_index].band[level][1].qlog;
+                else                    q= get_symbol(&s->c, s->header_state, 1);
+                s->plane[plane_index].band[level][orientation].qlog= q;
+            }
         }
-        pix += line_size - w;
     }
-    return s;
 }
 
-static inline void set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int ref, int type){
-    const int w= s->b_width << s->block_max_depth;
-    const int rem_depth= s->block_max_depth - level;
-    const int index= (x + y*w) << rem_depth;
-    const int block_w= 1<<rem_depth;
-    BlockNode block;
-    int i,j;
+#define GET_S(dst, check) \
+    tmp= get_symbol(&s->c, s->header_state, 0);\
+    if(!(check)){\
+        av_log(s->avctx, AV_LOG_ERROR, "Error " #dst " is %d\n", tmp);\
+        return -1;\
+    }\
+    dst= tmp;
 
-    block.color[0]= l;
-    block.color[1]= cb;
-    block.color[2]= cr;
-    block.mx= mx;
-    block.my= my;
-    block.ref= ref;
-    block.type= type;
-    block.level= level;
+static int decode_header(SnowContext *s){
+    int plane_index, tmp;
+    uint8_t kstate[32];
 
-    for(j=0; j<block_w; j++){
-        for(i=0; i<block_w; i++){
-            s->block[index + i + j*w]= block;
+    memset(kstate, MID_STATE, sizeof(kstate));
+
+    s->keyframe= get_rac(&s->c, kstate);
+    if(s->keyframe || s->always_reset){
+        reset_contexts(s);
+        s->spatial_decomposition_type=
+        s->qlog=
+        s->qbias=
+        s->mv_scale=
+        s->block_max_depth= 0;
+    }
+    if(s->keyframe){
+        GET_S(s->version, tmp <= 0U)
+        s->always_reset= get_rac(&s->c, s->header_state);
+        s->temporal_decomposition_type= get_symbol(&s->c, s->header_state, 0);
+        s->temporal_decomposition_count= get_symbol(&s->c, s->header_state, 0);
+        GET_S(s->spatial_decomposition_count, 0 < tmp && tmp <= MAX_DECOMPOSITIONS)
+        s->colorspace_type= get_symbol(&s->c, s->header_state, 0);
+        s->chroma_h_shift= get_symbol(&s->c, s->header_state, 0);
+        s->chroma_v_shift= get_symbol(&s->c, s->header_state, 0);
+        s->spatial_scalability= get_rac(&s->c, s->header_state);
+//        s->rate_scalability= get_rac(&s->c, s->header_state);
+        GET_S(s->max_ref_frames, tmp < (unsigned)MAX_REF_FRAMES)
+        s->max_ref_frames++;
+
+        decode_qlogs(s);
+    }
+
+    if(!s->keyframe){
+        if(get_rac(&s->c, s->header_state)){
+            for(plane_index=0; plane_index<2; plane_index++){
+                int htaps, i, sum=0;
+                Plane *p= &s->plane[plane_index];
+                p->diag_mc= get_rac(&s->c, s->header_state);
+                htaps= get_symbol(&s->c, s->header_state, 0)*2 + 2;
+                if((unsigned)htaps > HTAPS_MAX || htaps==0)
+                    return -1;
+                p->htaps= htaps;
+                for(i= htaps/2; i; i--){
+                    p->hcoeff[i]= get_symbol(&s->c, s->header_state, 0) * (1-2*(i&1));
+                    sum += p->hcoeff[i];
+                }
+                p->hcoeff[0]= 32-sum;
+            }
+            s->plane[2].diag_mc= s->plane[1].diag_mc;
+            s->plane[2].htaps  = s->plane[1].htaps;
+            memcpy(s->plane[2].hcoeff, s->plane[1].hcoeff, sizeof(s->plane[1].hcoeff));
+        }
+        if(get_rac(&s->c, s->header_state)){
+            GET_S(s->spatial_decomposition_count, 0 < tmp && tmp <= MAX_DECOMPOSITIONS)
+            decode_qlogs(s);
         }
     }
-}
 
-static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
-    const int offset[3]= {
-          y*c->  stride + x,
-        ((y*c->uvstride + x)>>1),
-        ((y*c->uvstride + x)>>1),
-    };
-    int i;
-    for(i=0; i<3; i++){
-        c->src[0][i]= src [i];
-        c->ref[0][i]= ref [i] + offset[i];
+    s->spatial_decomposition_type+= get_symbol(&s->c, s->header_state, 1);
+    if(s->spatial_decomposition_type > 1U){
+        av_log(s->avctx, AV_LOG_ERROR, "spatial_decomposition_type %d not supported", s->spatial_decomposition_type);
+        return -1;
     }
-    assert(!ref_index);
+    if(FFMIN(s->avctx-> width>>s->chroma_h_shift,
+             s->avctx->height>>s->chroma_v_shift) >> (s->spatial_decomposition_count-1) <= 0){
+        av_log(s->avctx, AV_LOG_ERROR, "spatial_decomposition_count %d too large for size", s->spatial_decomposition_count);
+        return -1;
+    }
+
+    s->qlog           += get_symbol(&s->c, s->header_state, 1);
+    s->mv_scale       += get_symbol(&s->c, s->header_state, 1);
+    s->qbias          += get_symbol(&s->c, s->header_state, 1);
+    s->block_max_depth+= get_symbol(&s->c, s->header_state, 1);
+    if(s->block_max_depth > 1 || s->block_max_depth < 0){
+        av_log(s->avctx, AV_LOG_ERROR, "block_max_depth= %d is too large", s->block_max_depth);
+        s->block_max_depth= 0;
+        return -1;
+    }
+
+    return 0;
 }
 
-static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
-                           const BlockNode *left, const BlockNode *top, const BlockNode *tr){
-    if(s->ref_frames == 1){
-        *mx = mid_pred(left->mx, top->mx, tr->mx);
-        *my = mid_pred(left->my, top->my, tr->my);
-    }else{
-        const int *scale = scale_mv_ref[ref];
-        *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
-                       (top ->mx * scale[top ->ref] + 128) >>8,
-                       (tr  ->mx * scale[tr  ->ref] + 128) >>8);
-        *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
-                       (top ->my * scale[top ->ref] + 128) >>8,
-                       (tr  ->my * scale[tr  ->ref] + 128) >>8);
+static void init_qexp(void){
+    int i;
+    double v=128;
+
+    for(i=0; i<QROOT; i++){
+        qexp[i]= lrintf(v);
+        v *= pow(2, 1.0 / QROOT);
     }
 }
 
-//FIXME copy&paste
-#define P_LEFT P[1]
-#define P_TOP P[2]
-#define P_TOPRIGHT P[3]
-#define P_MEDIAN P[4]
-#define P_MV1 P[9]
-#define FLAG_QPEL   1 //must be 1
+static av_cold int common_init(AVCodecContext *avctx){
+    SnowContext *s = avctx->priv_data;
+    int width, height;
+    int i, j;
 
-static int encode_q_branch(SnowContext *s, int level, int x, int y){
-    uint8_t p_buffer[1024];
-    uint8_t i_buffer[1024];
-    uint8_t p_state[sizeof(s->block_state)];
-    uint8_t i_state[sizeof(s->block_state)];
-    RangeCoder pc, ic;
-    uint8_t *pbbak= s->c.bytestream;
-    uint8_t *pbbak_start= s->c.bytestream_start;
-    int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
-    const int w= s->b_width  << s->block_max_depth;
-    const int h= s->b_height << s->block_max_depth;
-    const int rem_depth= s->block_max_depth - level;
-    const int index= (x + y*w) << rem_depth;
-    const int block_w= 1<<(LOG2_MB_SIZE - level);
-    int trx= (x+1)<<rem_depth;
-    int try= (y+1)<<rem_depth;
-    const BlockNode *left  = x ? &s->block[index-1] : &null_block;
-    const BlockNode *top   = y ? &s->block[index-w] : &null_block;
-    const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
-    const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
-    const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
-    const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
-    int pl = left->color[0];
-    int pcb= left->color[1];
-    int pcr= left->color[2];
-    int pmx, pmy;
-    int mx=0, my=0;
-    int l,cr,cb;
-    const int stride= s->current_picture.linesize[0];
-    const int uvstride= s->current_picture.linesize[1];
-    uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y*  stride)*block_w,
-                                s->input_picture.data[1] + (x + y*uvstride)*block_w/2,
-                                s->input_picture.data[2] + (x + y*uvstride)*block_w/2};
-    int P[10][2];
-    int16_t last_mv[3][2];
-    int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
-    const int shift= 1+qpel;
-    MotionEstContext *c= &s->m.me;
-    int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
-    int mx_context= av_log2(2*FFABS(left->mx - top->mx));
-    int my_context= av_log2(2*FFABS(left->my - top->my));
-    int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
-    int ref, best_ref, ref_score, ref_mx, ref_my;
+    s->avctx= avctx;
+    s->max_ref_frames=1; //just make sure its not an invalid value in case of no initial keyframe
 
-    assert(sizeof(s->block_state) >= 256);
-    if(s->keyframe){
-        set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
-        return 0;
-    }
+    dsputil_init(&s->dsp, avctx);
+    ff_dwt_init(&s->dwt);
 
-//    clip predictors / edge ?
+#define mcf(dx,dy)\
+    s->dsp.put_qpel_pixels_tab       [0][dy+dx/4]=\
+    s->dsp.put_no_rnd_qpel_pixels_tab[0][dy+dx/4]=\
+        s->dsp.put_h264_qpel_pixels_tab[0][dy+dx/4];\
+    s->dsp.put_qpel_pixels_tab       [1][dy+dx/4]=\
+    s->dsp.put_no_rnd_qpel_pixels_tab[1][dy+dx/4]=\
+        s->dsp.put_h264_qpel_pixels_tab[1][dy+dx/4];
 
-    P_LEFT[0]= left->mx;
-    P_LEFT[1]= left->my;
-    P_TOP [0]= top->mx;
-    P_TOP [1]= top->my;
-    P_TOPRIGHT[0]= tr->mx;
-    P_TOPRIGHT[1]= tr->my;
+    mcf( 0, 0)
+    mcf( 4, 0)
+    mcf( 8, 0)
+    mcf(12, 0)
+    mcf( 0, 4)
+    mcf( 4, 4)
+    mcf( 8, 4)
+    mcf(12, 4)
+    mcf( 0, 8)
+    mcf( 4, 8)
+    mcf( 8, 8)
+    mcf(12, 8)
+    mcf( 0,12)
+    mcf( 4,12)
+    mcf( 8,12)
+    mcf(12,12)
 
-    last_mv[0][0]= s->block[index].mx;
-    last_mv[0][1]= s->block[index].my;
-    last_mv[1][0]= right->mx;
-    last_mv[1][1]= right->my;
-    last_mv[2][0]= bottom->mx;
-    last_mv[2][1]= bottom->my;
+#define mcfh(dx,dy)\
+    s->dsp.put_pixels_tab       [0][dy/4+dx/8]=\
+    s->dsp.put_no_rnd_pixels_tab[0][dy/4+dx/8]=\
+        mc_block_hpel ## dx ## dy ## 16;\
+    s->dsp.put_pixels_tab       [1][dy/4+dx/8]=\
+    s->dsp.put_no_rnd_pixels_tab[1][dy/4+dx/8]=\
+        mc_block_hpel ## dx ## dy ## 8;
 
-    s->m.mb_stride=2;
-    s->m.mb_x=
-    s->m.mb_y= 0;
-    c->skip= 0;
+    mcfh(0, 0)
+    mcfh(8, 0)
+    mcfh(0, 8)
+    mcfh(8, 8)
 
-    assert(c->  stride ==   stride);
-    assert(c->uvstride == uvstride);
+    if(!qexp[0])
+        init_qexp();
 
-    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
-    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
-    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
-    c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
+//    dec += FFMAX(s->chroma_h_shift, s->chroma_v_shift);
 
-    c->xmin = - x*block_w - 16+2;
-    c->ymin = - y*block_w - 16+2;
-    c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-2;
-    c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-2;
+    width= s->avctx->width;
+    height= s->avctx->height;
 
-    if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
-    if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
-    if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
-    if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
-    if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
-    if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
-    if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
+    s->spatial_idwt_buffer= av_mallocz(width*height*sizeof(IDWTELEM));
+    s->spatial_dwt_buffer= av_mallocz(width*height*sizeof(DWTELEM)); //FIXME this does not belong here
 
-    P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
-    P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
+    for(i=0; i<MAX_REF_FRAMES; i++)
+        for(j=0; j<MAX_REF_FRAMES; j++)
+            scale_mv_ref[i][j] = 256*(i+1)/(j+1);
 
-    if (!y) {
-        c->pred_x= P_LEFT[0];
-        c->pred_y= P_LEFT[1];
-    } else {
-        c->pred_x = P_MEDIAN[0];
-        c->pred_y = P_MEDIAN[1];
-    }
+    s->avctx->get_buffer(s->avctx, &s->mconly_picture);
+    s->scratchbuf = av_malloc(s->mconly_picture.linesize[0]*7*MB_SIZE);
 
-    score= INT_MAX;
-    best_ref= 0;
-    for(ref=0; ref<s->ref_frames; ref++){
-        init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
+    return 0;
+}
 
-        ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
-                                         (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
+static int common_init_after_header(AVCodecContext *avctx){
+    SnowContext *s = avctx->priv_data;
+    int plane_index, level, orientation;
 
-        assert(ref_mx >= c->xmin);
-        assert(ref_mx <= c->xmax);
-        assert(ref_my >= c->ymin);
-        assert(ref_my <= c->ymax);
+    for(plane_index=0; plane_index<3; plane_index++){
+        int w= s->avctx->width;
+        int h= s->avctx->height;
 
-        ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
-        ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
-        ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
-        if(s->ref_mvs[ref]){
-            s->ref_mvs[ref][index][0]= ref_mx;
-            s->ref_mvs[ref][index][1]= ref_my;
-            s->ref_scores[ref][index]= ref_score;
-        }
-        if(score > ref_score){
-            score= ref_score;
-            best_ref= ref;
-            mx= ref_mx;
-            my= ref_my;
+        if(plane_index){
+            w>>= s->chroma_h_shift;
+            h>>= s->chroma_v_shift;
         }
-    }
-    //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
+        s->plane[plane_index].width = w;
+        s->plane[plane_index].height= h;
 
-  //  subpel search
-    base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
-    pc= s->c;
-    pc.bytestream_start=
-    pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
-    memcpy(p_state, s->block_state, sizeof(s->block_state));
+        for(level=s->spatial_decomposition_count-1; level>=0; level--){
+            for(orientation=level ? 1 : 0; orientation<4; orientation++){
+                SubBand *b= &s->plane[plane_index].band[level][orientation];
 
-    if(level!=s->block_max_depth)
-        put_rac(&pc, &p_state[4 + s_context], 1);
-    put_rac(&pc, &p_state[1 + left->type + top->type], 0);
-    if(s->ref_frames > 1)
-        put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
-    pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
-    put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
-    put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
-    p_len= pc.bytestream - pc.bytestream_start;
-    score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
+                b->buf= s->spatial_dwt_buffer;
+                b->level= level;
+                b->stride= s->plane[plane_index].width << (s->spatial_decomposition_count - level);
+                b->width = (w + !(orientation&1))>>1;
+                b->height= (h + !(orientation>1))>>1;
 
-    block_s= block_w*block_w;
-    sum = pix_sum(current_data[0], stride, block_w);
-    l= (sum + block_s/2)/block_s;
-    iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
+                b->stride_line = 1 << (s->spatial_decomposition_count - level);
+                b->buf_x_offset = 0;
+                b->buf_y_offset = 0;
 
-    block_s= block_w*block_w>>2;
-    sum = pix_sum(current_data[1], uvstride, block_w>>1);
-    cb= (sum + block_s/2)/block_s;
-//    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
-    sum = pix_sum(current_data[2], uvstride, block_w>>1);
-    cr= (sum + block_s/2)/block_s;
-//    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
+                if(orientation&1){
+                    b->buf += (w+1)>>1;
+                    b->buf_x_offset = (w+1)>>1;
+                }
+                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);
 
-    ic= s->c;
-    ic.bytestream_start=
-    ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
-    memcpy(i_state, s->block_state, sizeof(s->block_state));
-    if(level!=s->block_max_depth)
-        put_rac(&ic, &i_state[4 + s_context], 1);
-    put_rac(&ic, &i_state[1 + left->type + top->type], 1);
-    put_symbol(&ic, &i_state[32],  l-pl , 1);
-    put_symbol(&ic, &i_state[64], cb-pcb, 1);
-    put_symbol(&ic, &i_state[96], cr-pcr, 1);
-    i_len= ic.bytestream - ic.bytestream_start;
-    iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
+                if(level)
+                    b->parent= &s->plane[plane_index].band[level-1][orientation];
+                //FIXME avoid this realloc
+                av_freep(&b->x_coeff);
+                b->x_coeff=av_mallocz(((b->width+1) * b->height+1)*sizeof(x_and_coeff));
+            }
+            w= (w+1)>>1;
+            h= (h+1)>>1;
+        }
+    }
 
-//    assert(score==256*256*256*64-1);
-    assert(iscore < 255*255*256 + s->lambda2*10);
-    assert(iscore >= 0);
-    assert(l>=0 && l<=255);
-    assert(pl>=0 && pl<=255);
+    return 0;
+}
 
-    if(level==0){
-        int varc= iscore >> 8;
-        int vard= score >> 8;
-        if (vard <= 64 || vard < varc)
-            c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
-        else
-            c->scene_change_score+= s->m.qscale;
-    }
+#define QUANTIZE2 0
 
-    if(level!=s->block_max_depth){
-        put_rac(&s->c, &s->block_state[4 + s_context], 0);
-        score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
-        score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
-        score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
-        score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
-        score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
+#if QUANTIZE2==1
+#define Q2_STEP 8
 
-        if(score2 < score && score2 < iscore)
-            return score2;
-    }
+static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
+    SubBand *b= &p->band[level][orientation];
+    int x, y;
+    int xo=0;
+    int yo=0;
+    int step= 1 << (s->spatial_decomposition_count - level);
 
-    if(iscore < score){
-        pred_mv(s, &pmx, &pmy, 0, left, top, tr);
-        memcpy(pbbak, i_buffer, i_len);
-        s->c= ic;
-        s->c.bytestream_start= pbbak_start;
-        s->c.bytestream= pbbak + i_len;
-        set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
-        memcpy(s->block_state, i_state, sizeof(s->block_state));
-        return iscore;
-    }else{
-        memcpy(pbbak, p_buffer, p_len);
-        s->c= pc;
-        s->c.bytestream_start= pbbak_start;
-        s->c.bytestream= pbbak + p_len;
-        set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
-        memcpy(s->block_state, p_state, sizeof(s->block_state));
-        return score;
-    }
-}
+    if(orientation&1)
+        xo= step>>1;
+    if(orientation&2)
+        yo= step>>1;
 
-static av_always_inline int same_block(BlockNode *a, BlockNode *b){
-    if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
-        return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
-    }else{
-        return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
+    //FIXME bias for nonzero ?
+    //FIXME optimize
+    memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
+    for(y=0; y<p->height; y++){
+        for(x=0; x<p->width; x++){
+            int sx= (x-xo + step/2) / step / Q2_STEP;
+            int sy= (y-yo + step/2) / step / Q2_STEP;
+            int v= r0[x + y*p->width] - r1[x + y*p->width];
+            assert(sx>=0 && sy>=0 && sx < score_stride);
+            v= ((v+8)>>4)<<4;
+            score[sx + sy*score_stride] += v*v;
+            assert(score[sx + sy*score_stride] >= 0);
+        }
     }
 }
 
-static void encode_q_branch2(SnowContext *s, int level, int x, int y){
-    const int w= s->b_width  << s->block_max_depth;
-    const int rem_depth= s->block_max_depth - level;
-    const int index= (x + y*w) << rem_depth;
-    int trx= (x+1)<<rem_depth;
-    BlockNode *b= &s->block[index];
-    const BlockNode *left  = x ? &s->block[index-1] : &null_block;
-    const BlockNode *top   = y ? &s->block[index-w] : &null_block;
-    const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
-    const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
-    int pl = left->color[0];
-    int pcb= left->color[1];
-    int pcr= left->color[2];
-    int pmx, pmy;
-    int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
-    int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
-    int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
-    int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
+static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
+    int level, orientation;
 
-    if(s->keyframe){
-        set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
-        return;
-    }
+    for(level=0; level<s->spatial_decomposition_count; level++){
+        for(orientation=level ? 1 : 0; orientation<4; orientation++){
+            SubBand *b= &p->band[level][orientation];
+            IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
 
-    if(level!=s->block_max_depth){
-        if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
-            put_rac(&s->c, &s->block_state[4 + s_context], 1);
-        }else{
-            put_rac(&s->c, &s->block_state[4 + s_context], 0);
-            encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
-            encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
-            encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
-            encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
-            return;
+            dequantize(s, b, dst, b->stride);
         }
     }
-    if(b->type & BLOCK_INTRA){
-        pred_mv(s, &pmx, &pmy, 0, left, top, tr);
-        put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
-        put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
-        put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
-        put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
-        set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
-    }else{
-        pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
-        put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
-        if(s->ref_frames > 1)
-            put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
-        put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
-        put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
-        set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
-    }
 }
 
-static void decode_q_branch(SnowContext *s, int level, int x, int y){
-    const int w= s->b_width << s->block_max_depth;
-    const int rem_depth= s->block_max_depth - level;
-    const int index= (x + y*w) << rem_depth;
-    int trx= (x+1)<<rem_depth;
-    const BlockNode *left  = x ? &s->block[index-1] : &null_block;
-    const BlockNode *top   = y ? &s->block[index-w] : &null_block;
-    const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
-    const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
-    int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
+static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
+    int level, orientation, ys, xs, x, y, pass;
+    IDWTELEM best_dequant[height * stride];
+    IDWTELEM idwt2_buffer[height * stride];
+    const int score_stride= (width + 10)/Q2_STEP;
+    int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
+    int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
+    int threshold= (s->m.lambda * s->m.lambda) >> 6;
 
-    if(s->keyframe){
-        set_blocks(s, level, x, y, null_block.color[0], null_block.color[1], null_block.color[2], null_block.mx, null_block.my, null_block.ref, BLOCK_INTRA);
-        return;
-    }
+    //FIXME pass the copy cleanly ?
 
-    if(level==s->block_max_depth || get_rac(&s->c, &s->block_state[4 + s_context])){
-        int type, mx, my;
-        int l = left->color[0];
-        int cb= left->color[1];
-        int cr= left->color[2];
-        int ref = 0;
-        int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
-        int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 0*av_log2(2*FFABS(tr->mx - top->mx));
-        int my_context= av_log2(2*FFABS(left->my - top->my)) + 0*av_log2(2*FFABS(tr->my - top->my));
+//    memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
+    ff_spatial_dwt(buffer, width, height, stride, type, s->spatial_decomposition_count);
 
-        type= get_rac(&s->c, &s->block_state[1 + left->type + top->type]) ? BLOCK_INTRA : 0;
+    for(level=0; level<s->spatial_decomposition_count; level++){
+        for(orientation=level ? 1 : 0; orientation<4; orientation++){
+            SubBand *b= &p->band[level][orientation];
+            IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
+             DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
+            assert(src == b->buf); // code does not depend on this but it is true currently
 
-        if(type){
-            pred_mv(s, &mx, &my, 0, left, top, tr);
-            l += get_symbol(&s->c, &s->block_state[32], 1);
-            cb+= get_symbol(&s->c, &s->block_state[64], 1);
-            cr+= get_symbol(&s->c, &s->block_state[96], 1);
-        }else{
-            if(s->ref_frames > 1)
-                ref= get_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], 0);
-            pred_mv(s, &mx, &my, ref, left, top, tr);
-            mx+= get_symbol(&s->c, &s->block_state[128 + 32*(mx_context + 16*!!ref)], 1);
-            my+= get_symbol(&s->c, &s->block_state[128 + 32*(my_context + 16*!!ref)], 1);
+            quantize(s, b, dst, src, b->stride, s->qbias);
         }
-        set_blocks(s, level, x, y, l, cb, cr, mx, my, ref, type);
-    }else{
-        decode_q_branch(s, level+1, 2*x+0, 2*y+0);
-        decode_q_branch(s, level+1, 2*x+1, 2*y+0);
-        decode_q_branch(s, level+1, 2*x+0, 2*y+1);
-        decode_q_branch(s, level+1, 2*x+1, 2*y+1);
     }
-}
-
-static void encode_blocks(SnowContext *s, int search){
-    int x, y;
-    int w= s->b_width;
-    int h= s->b_height;
-
-    if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
-        iterative_me(s);
+    for(pass=0; pass<1; pass++){
+        if(s->qbias == 0) //keyframe
+            continue;
+        for(level=0; level<s->spatial_decomposition_count; level++){
+            for(orientation=level ? 1 : 0; orientation<4; orientation++){
+                SubBand *b= &p->band[level][orientation];
+                IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
+                IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
 
-    for(y=0; y<h; y++){
-        if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
-            av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
-            return;
-        }
-        for(x=0; x<w; x++){
-            if(s->avctx->me_method == ME_ITER || !search)
-                encode_q_branch2(s, 0, x, y);
-            else
-                encode_q_branch (s, 0, x, y);
+                for(ys= 0; ys<Q2_STEP; ys++){
+                    for(xs= 0; xs<Q2_STEP; xs++){
+                        memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
+                        dequantize_all(s, p, idwt2_buffer, width, height);
+                        ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
+                        find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
+                        memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
+                        for(y=ys; y<b->height; y+= Q2_STEP){
+                            for(x=xs; x<b->width; x+= Q2_STEP){
+                                if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
+                                if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
+                                //FIXME try more than just --
+                            }
+                        }
+                        dequantize_all(s, p, idwt2_buffer, width, height);
+                        ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
+                        find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
+                        for(y=ys; y<b->height; y+= Q2_STEP){
+                            for(x=xs; x<b->width; x+= Q2_STEP){
+                                int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
+                                if(score[score_idx] <= best_score[score_idx] + threshold){
+                                    best_score[score_idx]= score[score_idx];
+                                    if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
+                                    if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
+                                    //FIXME copy instead
+                                }
+                            }
+                        }
+                    }
+                }
+            }
         }
     }
+    memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
 }
 
-static void decode_blocks(SnowContext *s){
-    int x, y;
-    int w= s->b_width;
-    int h= s->b_height;
+#endif /* QUANTIZE2==1 */
 
-    for(y=0; y<h; y++){
-        for(x=0; x<w; x++){
-            decode_q_branch(s, 0, x, y);
-        }
-    }
-}
+#define USE_HALFPEL_PLANE 0
 
-static void mc_block(Plane *p, uint8_t *dst, const uint8_t *src, uint8_t *tmp, int stride, int b_w, int b_h, int dx, int dy){
-    static const uint8_t weight[64]={
-    8,7,6,5,4,3,2,1,
-    7,7,0,0,0,0,0,1,
-    6,0,6,0,0,0,2,0,
-    5,0,0,5,0,3,0,0,
-    4,0,0,0,4,0,0,0,
-    3,0,0,5,0,3,0,0,
-    2,0,6,0,0,0,2,0,
-    1,7,0,0,0,0,0,1,
-    };
+static void halfpel_interpol(SnowContext *s, uint8_t *halfpel[4][4], AVFrame *frame){
+    int p,x,y;
 
-    static const uint8_t brane[256]={
-    0x00,0x01,0x01,0x01,0x01,0x01,0x01,0x01,0x11,0x12,0x12,0x12,0x12,0x12,0x12,0x12,
-    0x04,0x05,0xcc,0xcc,0xcc,0xcc,0xcc,0x41,0x15,0x16,0xcc,0xcc,0xcc,0xcc,0xcc,0x52,
-    0x04,0xcc,0x05,0xcc,0xcc,0xcc,0x41,0xcc,0x15,0xcc,0x16,0xcc,0xcc,0xcc,0x52,0xcc,
-    0x04,0xcc,0xcc,0x05,0xcc,0x41,0xcc,0xcc,0x15,0xcc,0xcc,0x16,0xcc,0x52,0xcc,0xcc,
-    0x04,0xcc,0xcc,0xcc,0x41,0xcc,0xcc,0xcc,0x15,0xcc,0xcc,0xcc,0x16,0xcc,0xcc,0xcc,
-    0x04,0xcc,0xcc,0x41,0xcc,0x05,0xcc,0xcc,0x15,0xcc,0xcc,0x52,0xcc,0x16,0xcc,0xcc,
-    0x04,0xcc,0x41,0xcc,0xcc,0xcc,0x05,0xcc,0x15,0xcc,0x52,0xcc,0xcc,0xcc,0x16,0xcc,
-    0x04,0x41,0xcc,0xcc,0xcc,0xcc,0xcc,0x05,0x15,0x52,0xcc,0xcc,0xcc,0xcc,0xcc,0x16,
-    0x44,0x45,0x45,0x45,0x45,0x45,0x45,0x45,0x55,0x56,0x56,0x56,0x56,0x56,0x56,0x56,
-    0x48,0x49,0xcc,0xcc,0xcc,0xcc,0xcc,0x85,0x59,0x5A,0xcc,0xcc,0xcc,0xcc,0xcc,0x96,
-    0x48,0xcc,0x49,0xcc,0xcc,0xcc,0x85,0xcc,0x59,0xcc,0x5A,0xcc,0xcc,0xcc,0x96,0xcc,
-    0x48,0xcc,0xcc,0x49,0xcc,0x85,0xcc,0xcc,0x59,0xcc,0xcc,0x5A,0xcc,0x96,0xcc,0xcc,
-    0x48,0xcc,0xcc,0xcc,0x49,0xcc,0xcc,0xcc,0x59,0xcc,0xcc,0xcc,0x96,0xcc,0xcc,0xcc,
-    0x48,0xcc,0xcc,0x85,0xcc,0x49,0xcc,0xcc,0x59,0xcc,0xcc,0x96,0xcc,0x5A,0xcc,0xcc,
-    0x48,0xcc,0x85,0xcc,0xcc,0xcc,0x49,0xcc,0x59,0xcc,0x96,0xcc,0xcc,0xcc,0x5A,0xcc,
-    0x48,0x85,0xcc,0xcc,0xcc,0xcc,0xcc,0x49,0x59,0x96,0xcc,0xcc,0xcc,0xcc,0xcc,0x5A,
-    };
-
-    static const uint8_t needs[16]={
-    0,1,0,0,
-    2,4,2,0,
-    0,1,0,0,
-    15
-    };
+    assert(!(s->avctx->flags & CODEC_FLAG_EMU_EDGE));
 
-    int x, y, b, r, l;
-    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;
+    for(p=0; p<3; p++){
+        int is_chroma= !!p;
+        int w= s->avctx->width  >>is_chroma;
+        int h= s->avctx->height >>is_chroma;
+        int ls= frame->linesize[p];
+        uint8_t *src= frame->data[p];
 
-    b= needs[l] | needs[r];
-    if(p && !p->diag_mc)
-        b= 15;
+        halfpel[1][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls);
+        halfpel[2][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls);
+        halfpel[3][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls);
 
-    if(b&5){
-        for(y=0; y < b_h+HTAPS_MAX-1; y++){
-            for(x=0; x < b_w; x++){
-                int a_1=src[x + HTAPS_MAX/2-4];
-                int a0= src[x + HTAPS_MAX/2-3];
-                int a1= src[x + HTAPS_MAX/2-2];
-                int a2= src[x + HTAPS_MAX/2-1];
-                int a3= src[x + HTAPS_MAX/2+0];
-                int a4= src[x + HTAPS_MAX/2+1];
-                int a5= src[x + HTAPS_MAX/2+2];
-                int a6= src[x + HTAPS_MAX/2+3];
-                int am=0;
-                if(!p || p->fast_mc){
-                    am= 20*(a2+a3) - 5*(a1+a4) + (a0+a5);
-                    tmpI[x]= am;
-                    am= (am+16)>>5;
-                }else{
-                    am= p->hcoeff[0]*(a2+a3) + p->hcoeff[1]*(a1+a4) + p->hcoeff[2]*(a0+a5) + p->hcoeff[3]*(a_1+a6);
-                    tmpI[x]= am;
-                    am= (am+32)>>6;
-                }
+        halfpel[0][p]= src;
+        for(y=0; y<h; y++){
+            for(x=0; x<w; x++){
+                int i= y*ls + x;
 
-                if(am&(~255)) am= ~(am>>31);
-                tmp2[x]= am;
+                halfpel[1][p][i]= (20*(src[i] + src[i+1]) - 5*(src[i-1] + src[i+2]) + (src[i-2] + src[i+3]) + 16 )>>5;
             }
-            tmpI+= 64;
-            tmp2+= stride;
-            src += stride;
         }
-        src -= stride*y;
-    }
-    src += HTAPS_MAX/2 - 1;
-    tmp2= tmp2t[1];
-
-    if(b&2){
-        for(y=0; y < b_h; y++){
-            for(x=0; x < b_w+1; x++){
-                int a_1=src[x + (HTAPS_MAX/2-4)*stride];
-                int a0= src[x + (HTAPS_MAX/2-3)*stride];
-                int a1= src[x + (HTAPS_MAX/2-2)*stride];
-                int a2= src[x + (HTAPS_MAX/2-1)*stride];
-                int a3= src[x + (HTAPS_MAX/2+0)*stride];
-                int a4= src[x + (HTAPS_MAX/2+1)*stride];
-                int a5= src[x + (HTAPS_MAX/2+2)*stride];
-                int a6= src[x + (HTAPS_MAX/2+3)*stride];
-                int am=0;
-                if(!p || p->fast_mc)
-                    am= (20*(a2+a3) - 5*(a1+a4) + (a0+a5) + 16)>>5;
-                else
-                    am= (p->hcoeff[0]*(a2+a3) + p->hcoeff[1]*(a1+a4) + p->hcoeff[2]*(a0+a5) + p->hcoeff[3]*(a_1+a6) + 32)>>6;
+        for(y=0; y<h; y++){
+            for(x=0; x<w; x++){
+                int i= y*ls + x;
 
-                if(am&(~255)) am= ~(am>>31);
-                tmp2[x]= am;
+                halfpel[2][p][i]= (20*(src[i] + src[i+ls]) - 5*(src[i-ls] + src[i+2*ls]) + (src[i-2*ls] + src[i+3*ls]) + 16 )>>5;
             }
-            src += stride;
-            tmp2+= stride;
         }
-        src -= stride*y;
-    }
-    src += stride*(HTAPS_MAX/2 - 1);
-    tmp2= tmp2t[2];
-    tmpI= tmpIt;
-    if(b&4){
-        for(y=0; y < b_h; y++){
-            for(x=0; x < b_w; x++){
-                int a_1=tmpI[x + (HTAPS_MAX/2-4)*64];
-                int a0= tmpI[x + (HTAPS_MAX/2-3)*64];
-                int a1= tmpI[x + (HTAPS_MAX/2-2)*64];
-                int a2= tmpI[x + (HTAPS_MAX/2-1)*64];
-                int a3= tmpI[x + (HTAPS_MAX/2+0)*64];
-                int a4= tmpI[x + (HTAPS_MAX/2+1)*64];
-                int a5= tmpI[x + (HTAPS_MAX/2+2)*64];
-                int a6= tmpI[x + (HTAPS_MAX/2+3)*64];
-                int am=0;
-                if(!p || p->fast_mc)
-                    am= (20*(a2+a3) - 5*(a1+a4) + (a0+a5) + 512)>>10;
-                else
-                    am= (p->hcoeff[0]*(a2+a3) + p->hcoeff[1]*(a1+a4) + p->hcoeff[2]*(a0+a5) + p->hcoeff[3]*(a_1+a6) + 2048)>>12;
-                if(am&(~255)) am= ~(am>>31);
-                tmp2[x]= am;
+        src= halfpel[1][p];
+        for(y=0; y<h; y++){
+            for(x=0; x<w; x++){
+                int i= y*ls + x;
+
+                halfpel[3][p][i]= (20*(src[i] + src[i+ls]) - 5*(src[i-ls] + src[i+2*ls]) + (src[i-2*ls] + src[i+3*ls]) + 16 )>>5;
             }
-            tmpI+= 64;
-            tmp2+= stride;
         }
+
+//FIXME border!
     }
+}
 
-    hpel[ 0]= src;
-    hpel[ 1]= tmp2t[0] + stride*(HTAPS_MAX/2-1);
-    hpel[ 2]= src + 1;
+static void release_buffer(AVCodecContext *avctx){
+    SnowContext *s = avctx->priv_data;
+    int i;
 
-    hpel[ 4]= tmp2t[1];
-    hpel[ 5]= tmp2t[2];
-    hpel[ 6]= tmp2t[1] + 1;
+    if(s->last_picture[s->max_ref_frames-1].data[0]){
+        avctx->release_buffer(avctx, &s->last_picture[s->max_ref_frames-1]);
+        for(i=0; i<9; i++)
+            if(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3])
+                av_free(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3] - EDGE_WIDTH*(1+s->current_picture.linesize[i%3]));
+    }
+}
 
-    hpel[ 8]= src + stride;
-    hpel[ 9]= hpel[1] + stride;
-    hpel[10]= hpel[8] + 1;
+static int frame_start(SnowContext *s){
+   AVFrame tmp;
+   int w= s->avctx->width; //FIXME round up to x16 ?
+   int h= s->avctx->height;
 
-    if(b==15){
-        const uint8_t *src1= hpel[dx/8 + dy/8*4  ];
-        const uint8_t *src2= hpel[dx/8 + dy/8*4+1];
-        const uint8_t *src3= hpel[dx/8 + dy/8*4+4];
-        const uint8_t *src4= hpel[dx/8 + dy/8*4+5];
-        dx&=7;
-        dy&=7;
-        for(y=0; y < b_h; y++){
-            for(x=0; x < b_w; x++){
-                dst[x]= ((8-dx)*(8-dy)*src1[x] + dx*(8-dy)*src2[x]+
-                         (8-dx)*   dy *src3[x] + dx*   dy *src4[x]+32)>>6;
-            }
-            src1+=stride;
-            src2+=stride;
-            src3+=stride;
-            src4+=stride;
-            dst +=stride;
-        }
+    if(s->current_picture.data[0]){
+        s->dsp.draw_edges(s->current_picture.data[0], s->current_picture.linesize[0], w   , h   , EDGE_WIDTH  , EDGE_TOP|EDGE_BOTTOM);
+        s->dsp.draw_edges(s->current_picture.data[1], s->current_picture.linesize[1], w>>1, h>>1, EDGE_WIDTH/2, EDGE_TOP|EDGE_BOTTOM);
+        s->dsp.draw_edges(s->current_picture.data[2], s->current_picture.linesize[2], w>>1, h>>1, EDGE_WIDTH/2, EDGE_TOP|EDGE_BOTTOM);
+    }
+
+    release_buffer(s->avctx);
+
+    tmp= s->last_picture[s->max_ref_frames-1];
+    memmove(s->last_picture+1, s->last_picture, (s->max_ref_frames-1)*sizeof(AVFrame));
+    memmove(s->halfpel_plane+1, s->halfpel_plane, (s->max_ref_frames-1)*sizeof(void*)*4*4);
+    if(USE_HALFPEL_PLANE && s->current_picture.data[0])
+        halfpel_interpol(s, s->halfpel_plane[0], &s->current_picture);
+    s->last_picture[0]= s->current_picture;
+    s->current_picture= tmp;
+
+    if(s->keyframe){
+        s->ref_frames= 0;
     }else{
-        const uint8_t *src1= hpel[l];
-        const uint8_t *src2= hpel[r];
-        int a= weight[((dx&7) + (8*(dy&7)))];
-        int b= 8-a;
-        for(y=0; y < b_h; y++){
-            for(x=0; x < b_w; x++){
-                dst[x]= (a*src1[x] + b*src2[x] + 4)>>3;
-            }
-            src1+=stride;
-            src2+=stride;
-            dst +=stride;
+        int i;
+        for(i=0; i<s->max_ref_frames && s->last_picture[i].data[0]; i++)
+            if(i && s->last_picture[i-1].key_frame)
+                break;
+        s->ref_frames= i;
+        if(s->ref_frames==0){
+            av_log(s->avctx,AV_LOG_ERROR, "No reference frames\n");
+            return -1;
         }
     }
-}
 
-#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)];\
-    assert(h==b_w);\
-    mc_block(NULL, dst, src-(HTAPS_MAX/2-1)-(HTAPS_MAX/2-1)*stride, tmp, stride, b_w, b_w, dx, dy);\
+    s->current_picture.reference= 1;
+    if(s->avctx->get_buffer(s->avctx, &s->current_picture) < 0){
+        av_log(s->avctx, AV_LOG_ERROR, "get_buffer() failed\n");
+        return -1;
+    }
+
+    s->current_picture.key_frame= s->keyframe;
+
+    return 0;
 }
 
-mca( 0, 0,16)
-mca( 8, 0,16)
-mca( 0, 8,16)
-mca( 8, 8,16)
-mca( 0, 0,8)
-mca( 8, 0,8)
-mca( 0, 8,8)
-mca( 8, 8,8)
+static av_cold void common_end(SnowContext *s){
+    int plane_index, level, orientation, i;
 
-static void pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, int stride, int sx, int sy, int b_w, int b_h, BlockNode *block, int plane_index, int w, int h){
-    if(block->type & BLOCK_INTRA){
-        int x, y;
-        const int color = block->color[plane_index];
-        const int color4= color*0x01010101;
-        if(b_w==32){
-            for(y=0; y < b_h; y++){
-                *(uint32_t*)&dst[0 + y*stride]= color4;
-                *(uint32_t*)&dst[4 + y*stride]= color4;
-                *(uint32_t*)&dst[8 + y*stride]= color4;
-                *(uint32_t*)&dst[12+ y*stride]= color4;
-                *(uint32_t*)&dst[16+ y*stride]= color4;
-                *(uint32_t*)&dst[20+ y*stride]= color4;
-                *(uint32_t*)&dst[24+ y*stride]= color4;
-                *(uint32_t*)&dst[28+ y*stride]= color4;
-            }
-        }else if(b_w==16){
-            for(y=0; y < b_h; y++){
-                *(uint32_t*)&dst[0 + y*stride]= color4;
-                *(uint32_t*)&dst[4 + y*stride]= color4;
-                *(uint32_t*)&dst[8 + y*stride]= color4;
-                *(uint32_t*)&dst[12+ y*stride]= color4;
-            }
-        }else if(b_w==8){
-            for(y=0; y < b_h; y++){
-                *(uint32_t*)&dst[0 + y*stride]= color4;
-                *(uint32_t*)&dst[4 + y*stride]= color4;
-            }
-        }else if(b_w==4){
-            for(y=0; y < b_h; y++){
-                *(uint32_t*)&dst[0 + y*stride]= color4;
-            }
-        }else{
-            for(y=0; y < b_h; y++){
-                for(x=0; x < b_w; x++){
-                    dst[x + y*stride]= color;
-                }
-            }
-        }
-    }else{
-        uint8_t *src= s->last_picture[block->ref].data[plane_index];
-        const int scale= plane_index ?  s->mv_scale : 2*s->mv_scale;
-        int mx= block->mx*scale;
-        int my= block->my*scale;
-        const int dx= mx&15;
-        const int dy= my&15;
-        const int tab_index= 3 - (b_w>>2) + (b_w>>4);
-        sx += (mx>>4) - (HTAPS_MAX/2-1);
-        sy += (my>>4) - (HTAPS_MAX/2-1);
-        src += sx + sy*stride;
-        if(   (unsigned)sx >= w - b_w - (HTAPS_MAX-2)
-           || (unsigned)sy >= h - b_h - (HTAPS_MAX-2)){
-            ff_emulated_edge_mc(tmp + MB_SIZE, src, stride, b_w+HTAPS_MAX-1, b_h+HTAPS_MAX-1, sx, sy, w, h);
-            src= tmp + MB_SIZE;
-        }
-//        assert(b_w == b_h || 2*b_w == b_h || b_w == 2*b_h);
-//        assert(!(b_w&(b_w-1)));
-        assert(b_w>1 && b_h>1);
-        assert((tab_index>=0 && tab_index<4) || b_w==32);
-        if((dx&3) || (dy&3) || !(b_w == b_h || 2*b_w == b_h || b_w == 2*b_h) || (b_w&(b_w-1)) || !s->plane[plane_index].fast_mc )
-            mc_block(&s->plane[plane_index], dst, src, tmp, stride, b_w, b_h, dx, dy);
-        else if(b_w==32){
-            int y;
-            for(y=0; y<b_h; y+=16){
-                s->dsp.put_h264_qpel_pixels_tab[0][dy+(dx>>2)](dst + y*stride, src + 3 + (y+3)*stride,stride);
-                s->dsp.put_h264_qpel_pixels_tab[0][dy+(dx>>2)](dst + 16 + y*stride, src + 19 + (y+3)*stride,stride);
-            }
-        }else if(b_w==b_h)
-            s->dsp.put_h264_qpel_pixels_tab[tab_index  ][dy+(dx>>2)](dst,src + 3 + 3*stride,stride);
-        else if(b_w==2*b_h){
-            s->dsp.put_h264_qpel_pixels_tab[tab_index+1][dy+(dx>>2)](dst    ,src + 3       + 3*stride,stride);
-            s->dsp.put_h264_qpel_pixels_tab[tab_index+1][dy+(dx>>2)](dst+b_h,src + 3 + b_h + 3*stride,stride);
-        }else{
-            assert(2*b_w==b_h);
-            s->dsp.put_h264_qpel_pixels_tab[tab_index  ][dy+(dx>>2)](dst           ,src + 3 + 3*stride           ,stride);
-            s->dsp.put_h264_qpel_pixels_tab[tab_index  ][dy+(dx>>2)](dst+b_w*stride,src + 3 + 3*stride+b_w*stride,stride);
-        }
+    av_freep(&s->spatial_dwt_buffer);
+    av_freep(&s->spatial_idwt_buffer);
+
+    s->m.me.temp= NULL;
+    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);
+
+    av_freep(&s->block);
+    av_freep(&s->scratchbuf);
+
+    for(i=0; i<MAX_REF_FRAMES; i++){
+        av_freep(&s->ref_mvs[i]);
+        av_freep(&s->ref_scores[i]);
+        if(s->last_picture[i].data[0])
+            s->avctx->release_buffer(s->avctx, &s->last_picture[i]);
     }
-}
 
-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);
-        const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
-        dst = slice_buffer_get_line(sb, src_y + y);
-        for(x=0; x<b_w; x++){
-            int v=   obmc1[x] * block[3][x + y*src_stride]
-                    +obmc2[x] * block[2][x + y*src_stride]
-                    +obmc3[x] * block[1][x + y*src_stride]
-                    +obmc4[x] * block[0][x + y*src_stride];
+    for(plane_index=0; plane_index<3; plane_index++){
+        for(level=s->spatial_decomposition_count-1; level>=0; level--){
+            for(orientation=level ? 1 : 0; orientation<4; orientation++){
+                SubBand *b= &s->plane[plane_index].band[level][orientation];
 
-            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;
-                if(v&(~255)) v= ~(v>>31);
-                dst8[x + y*src_stride] = v;
-            }else{
-                dst[x + src_x] -= v;
+                av_freep(&b->x_coeff);
             }
         }
     }
+    if (s->mconly_picture.data[0])
+        s->avctx->release_buffer(s->avctx, &s->mconly_picture);
+    if (s->current_picture.data[0])
+        s->avctx->release_buffer(s->avctx, &s->current_picture);
 }
 
-//FIXME name cleanup (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;
-    BlockNode *lb= lt+b_stride;
-    BlockNode *rb= lb+1;
-    uint8_t *block[4];
-    int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
-    uint8_t *tmp = s->scratchbuf;
-    uint8_t *ptmp;
-    int x,y;
+static av_cold int decode_init(AVCodecContext *avctx)
+{
+    avctx->pix_fmt= PIX_FMT_YUV420P;
 
-    if(b_x<0){
-        lt= rt;
-        lb= rb;
-    }else if(b_x + 1 >= b_width){
-        rt= lt;
-        rb= lb;
-    }
-    if(b_y<0){
-        lt= lb;
-        rt= rb;
-    }else if(b_y + 1 >= b_height){
-        lb= lt;
-        rb= rt;
-    }
+    common_init(avctx);
 
-    if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
-        obmc -= src_x;
-        b_w += src_x;
-        if(!sliced && !offset_dst)
-            dst -= src_x;
-        src_x=0;
-    }else if(src_x + b_w > w){
-        b_w = w - src_x;
-    }
-    if(src_y<0){
-        obmc -= src_y*obmc_stride;
-        b_h += src_y;
-        if(!sliced && !offset_dst)
-            dst -= src_y*dst_stride;
-        src_y=0;
-    }else if(src_y + b_h> h){
-        b_h = h - src_y;
-    }
+    return 0;
+}
 
-    if(b_w<=0 || b_h<=0) return;
+static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, AVPacket *avpkt){
+    const uint8_t *buf = avpkt->data;
+    int buf_size = avpkt->size;
+    SnowContext *s = avctx->priv_data;
+    RangeCoder * const c= &s->c;
+    int bytes_read;
+    AVFrame *picture = data;
+    int level, orientation, plane_index;
 
-    assert(src_stride > 2*MB_SIZE + 5);
+    ff_init_range_decoder(c, buf, buf_size);
+    ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
 
-    if(!sliced && offset_dst)
-        dst += src_x + src_y*dst_stride;
-    dst8+= src_x + src_y*src_stride;
-//    src += src_x + src_y*src_stride;
+    s->current_picture.pict_type= AV_PICTURE_TYPE_I; //FIXME I vs. P
+    if(decode_header(s)<0)
+        return -1;
+    common_init_after_header(avctx);
 
-    ptmp= tmp + 3*tmp_step;
-    block[0]= ptmp;
-    ptmp+=tmp_step;
-    pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
+    // realloc slice buffer for the case that spatial_decomposition_count changed
+    ff_slice_buffer_destroy(&s->sb);
+    ff_slice_buffer_init(&s->sb, s->plane[0].height, (MB_SIZE >> s->block_max_depth) + s->spatial_decomposition_count * 8 + 1, s->plane[0].width, s->spatial_idwt_buffer);
 
-    if(same_block(lt, rt)){
-        block[1]= block[0];
-    }else{
-        block[1]= ptmp;
-        ptmp+=tmp_step;
-        pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
+    for(plane_index=0; plane_index<3; plane_index++){
+        Plane *p= &s->plane[plane_index];
+        p->fast_mc= p->diag_mc && p->htaps==6 && p->hcoeff[0]==40
+                                              && p->hcoeff[1]==-10
+                                              && p->hcoeff[2]==2;
     }
 
-    if(same_block(lt, lb)){
-        block[2]= block[0];
-    }else if(same_block(rt, lb)){
-        block[2]= block[1];
-    }else{
-        block[2]= ptmp;
-        ptmp+=tmp_step;
-        pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
-    }
+    alloc_blocks(s);
 
-    if(same_block(lt, rb) ){
-        block[3]= block[0];
-    }else if(same_block(rt, rb)){
-        block[3]= block[1];
-    }else if(same_block(lb, rb)){
-        block[3]= block[2];
-    }else{
-        block[3]= ptmp;
-        pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
-    }
-#if 0
-    for(y=0; y<b_h; y++){
-        for(x=0; x<b_w; x++){
-            int v=   obmc [x + y*obmc_stride] * block[3][x + y*src_stride] * (256/OBMC_MAX);
-            if(add) dst[x + y*dst_stride] += v;
-            else    dst[x + y*dst_stride] -= v;
+    if(frame_start(s) < 0)
+        return -1;
+    //keyframe flag duplication mess FIXME
+    if(avctx->debug&FF_DEBUG_PICT_INFO)
+        av_log(avctx, AV_LOG_ERROR, "keyframe:%d qlog:%d\n", s->keyframe, s->qlog);
+
+    decode_blocks(s);
+
+    for(plane_index=0; plane_index<3; plane_index++){
+        Plane *p= &s->plane[plane_index];
+        int w= p->width;
+        int h= p->height;
+        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;
+                }
+            }
         }
-    }
-    for(y=0; y<b_h; y++){
-        uint8_t *obmc2= obmc + (obmc_stride>>1);
-        for(x=0; x<b_w; x++){
-            int v=   obmc2[x + y*obmc_stride] * block[2][x + y*src_stride] * (256/OBMC_MAX);
-            if(add) dst[x + y*dst_stride] += v;
-            else    dst[x + y*dst_stride] -= 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);
+            }
         }
-    }
-    for(y=0; y<b_h; y++){
-        uint8_t *obmc3= obmc + obmc_stride*(obmc_stride>>1);
-        for(x=0; x<b_w; x++){
-            int v=   obmc3[x + y*obmc_stride] * block[1][x + y*src_stride] * (256/OBMC_MAX);
-            if(add) dst[x + y*dst_stride] += v;
-            else    dst[x + y*dst_stride] -= v;
-        }
-    }
-    for(y=0; y<b_h; y++){
-        uint8_t *obmc3= obmc + obmc_stride*(obmc_stride>>1);
-        uint8_t *obmc4= obmc3+ (obmc_stride>>1);
-        for(x=0; x<b_w; x++){
-            int v=   obmc4[x + y*obmc_stride] * block[0][x + y*src_stride] * (256/OBMC_MAX);
-            if(add) dst[x + y*dst_stride] += v;
-            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);
-            const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
-            const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
-            for(x=0; x<b_w; x++){
-                int v=   obmc1[x] * block[3][x + y*src_stride]
-                        +obmc2[x] * block[2][x + y*src_stride]
-                        +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;
-                    if(v&(~255)) v= ~(v>>31);
-                    dst8[x + y*src_stride] = v;
-                }else{
-                    dst[x + y*dst_stride] -= v;
-                }
-            }
-        }
-    }
-#endif /* 0 */
-}
+        {
+        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;
+        DWTCompose cs[MAX_DECOMPOSITIONS];
+        int yd=0, yq=0;
+        int y;
+        int end_y;
 
-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;
-    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];
-    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;
+        ff_spatial_idwt_buffered_init(cs, &s->sb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count);
+        for(mb_y=0; mb_y<=mb_h; mb_y++){
 
-    if(s->keyframe || (s->avctx->debug&512)){
-        if(mb_y==mb_h)
-            return;
+            int slice_starty = block_w*mb_y;
+            int slice_h = block_w*(mb_y+1);
+            if (!(s->keyframe || s->avctx->debug&512)){
+                slice_starty = FFMAX(0, slice_starty - (block_w >> 1));
+                slice_h -= (block_w >> 1);
+            }
 
-        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;
-                    if(v&(~255)) v= ~(v>>31);
-                    dst8[x + y*ref_stride]= v;
+            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;
+                    int our_mb_start = mb_y;
+                    int our_mb_end = (mb_y + 1);
+                    const int extra= 3;
+                    start_y = (mb_y ? ((block_w * our_mb_start) >> (s->spatial_decomposition_count - level)) + s->spatial_decomposition_count - level + extra: 0);
+                    end_y = (((block_w * our_mb_end) >> (s->spatial_decomposition_count - level)) + s->spatial_decomposition_count - level + extra);
+                    if (!(s->keyframe || s->avctx->debug&512)){
+                        start_y = FFMAX(0, start_y - (block_w >> (1+s->spatial_decomposition_count - level)));
+                        end_y = FFMAX(0, end_y - (block_w >> (1+s->spatial_decomposition_count - level)));
+                    }
+                    start_y = FFMIN(b->height, start_y);
+                    end_y = FFMIN(b->height, end_y);
+
+                    if (start_y != end_y){
+                        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]);
+                    }
                 }
             }
-        }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;
+
+            for(; yd<slice_h; yd+=4){
+                ff_spatial_idwt_buffered_slice(&s->dwt, 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++){
+                        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)
+                ff_slice_buffer_release(&s->sb, y++);
         }
 
-        return;
-    }
+        ff_slice_buffer_flush(&s->sb);
+        }
 
-    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 mb_h= s->b_height << s->block_max_depth;
-    int x, y, mb_x;
-    int block_size = MB_SIZE >> s->block_max_depth;
-    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;
-    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;
+    emms_c();
 
-    if(s->keyframe || (s->avctx->debug&512)){
-        if(mb_y==mb_h)
-            return;
+    release_buffer(avctx);
 
-        if(add){
-            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
-                for(x=0; x<w; x++){
-                    int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
-                    v >>= FRAC_BITS;
-                    if(v&(~255)) v= ~(v>>31);
-                    dst8[x + y*ref_stride]= v;
-                }
-            }
-        }else{
-            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
-                for(x=0; x<w; x++){
-                    buf[x + y*w]-= 128<<FRAC_BITS;
-                }
-            }
-        }
+    if(!(s->avctx->debug&2048))
+        *picture= s->current_picture;
+    else
+        *picture= s->mconly_picture;
 
-        return;
-    }
+    *data_size = sizeof(AVFrame);
 
-    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);
-    }
-}
+    bytes_read= c->bytestream - c->bytestream_start;
+    if(bytes_read ==0) av_log(s->avctx, AV_LOG_ERROR, "error at end of frame\n"); //FIXME
 
-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);
+    return bytes_read;
 }
 
-static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
-    int i, x2, y2;
-    Plane *p= &s->plane[plane_index];
-    const int block_size = MB_SIZE >> s->block_max_depth;
-    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];
-    BlockNode backup= *b;
-    int ab=0;
-    int aa=0;
-
-    b->type|= BLOCK_INTRA;
-    b->color[plane_index]= 0;
-    memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
+static av_cold int decode_end(AVCodecContext *avctx)
+{
+    SnowContext *s = avctx->priv_data;
 
-    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;
-        int y= block_w*mb_y2 + block_w/2;
+    ff_slice_buffer_destroy(&s->sb);
 
-        add_yblock(s, 0, NULL, dst + ((i&1)+(i>>1)*obmc_stride)*block_w, NULL, obmc,
-                    x, y, block_w, block_w, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
+    common_end(s);
 
-        for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_w); y2++){
-            for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
-                int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_w*mb_y - block_w/2))*obmc_stride;
-                int obmc_v= obmc[index];
-                int d;
-                if(y<0) obmc_v += obmc[index + block_w*obmc_stride];
-                if(x<0) obmc_v += obmc[index + block_w];
-                if(y+block_w>h) obmc_v += obmc[index - block_w*obmc_stride];
-                if(x+block_w>w) obmc_v += obmc[index - block_w];
-                //FIXME precalculate this or simplify it somehow else
+    return 0;
+}
 
-                d = -dst[index] + (1<<(FRAC_BITS-1));
-                dst[index] = d;
-                ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
-                aa += obmc_v * obmc_v; //FIXME precalculate this
-            }
-        }
+AVCodec ff_snow_decoder = {
+    "snow",
+    AVMEDIA_TYPE_VIDEO,
+    CODEC_ID_SNOW,
+    sizeof(SnowContext),
+    decode_init,
+    NULL,
+    decode_end,
+    decode_frame,
+    CODEC_CAP_DR1 /*| CODEC_CAP_DRAW_HORIZ_BAND*/,
+    NULL,
+    .long_name = NULL_IF_CONFIG_SMALL("Snow"),
+};
+
+#if CONFIG_SNOW_ENCODER
+static av_cold int encode_init(AVCodecContext *avctx)
+{
+    SnowContext *s = avctx->priv_data;
+    int plane_index;
+
+    if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
+        av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
+               "Use vstrict=-2 / -strict -2 to use it anyway.\n");
+        return -1;
     }
-    *b= backup;
 
-    return av_clip(((ab<<LOG2_OBMC_MAX) + aa/2)/aa, 0, 255); //FIXME we should not need clipping
-}
+    if(avctx->prediction_method == DWT_97
+       && (avctx->flags & CODEC_FLAG_QSCALE)
+       && avctx->global_quality == 0){
+        av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
+        return -1;
+    }
 
-static inline int get_block_bits(SnowContext *s, int x, int y, int w){
-    const int b_stride = s->b_width << s->block_max_depth;
-    const int b_height = s->b_height<< s->block_max_depth;
-    int index= x + y*b_stride;
-    const BlockNode *b     = &s->block[index];
-    const BlockNode *left  = x ? &s->block[index-1] : &null_block;
-    const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
-    const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
-    const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
-    int dmx, dmy;
-//  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
-//  int my_context= av_log2(2*FFABS(left->my - top->my));
+    s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
 
-    if(x<0 || x>=b_stride || y>=b_height)
-        return 0;
-/*
-1            0      0
-01X          1-2    1
-001XX        3-6    2-3
-0001XXX      7-14   4-7
-00001XXXX   15-30   8-15
-*/
-//FIXME try accurate rate
-//FIXME intra and inter predictors if surrounding blocks are not the same type
-    if(b->type & BLOCK_INTRA){
-        return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
-                   + av_log2(2*FFABS(left->color[1] - b->color[1]))
-                   + av_log2(2*FFABS(left->color[2] - b->color[2])));
-    }else{
-        pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
-        dmx-= b->mx;
-        dmy-= b->my;
-        return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
-                    + av_log2(2*FFABS(dmy))
-                    + av_log2(2*b->ref));
+    s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
+    s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
+
+    for(plane_index=0; plane_index<3; plane_index++){
+        s->plane[plane_index].diag_mc= 1;
+        s->plane[plane_index].htaps= 6;
+        s->plane[plane_index].hcoeff[0]=  40;
+        s->plane[plane_index].hcoeff[1]= -10;
+        s->plane[plane_index].hcoeff[2]=   2;
+        s->plane[plane_index].fast_mc= 1;
     }
-}
 
-static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, const uint8_t *obmc_edged){
-    Plane *p= &s->plane[plane_index];
-    const int block_size = MB_SIZE >> s->block_max_depth;
-    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 = s->scratchbuf;
-    uint8_t tmp[ref_stride*(2*MB_SIZE+HTAPS_MAX-1)];
-    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;
-    const int h= p->height;
-    int distortion;
-    int rate= 0;
-    const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
-    int sx= block_w*mb_x - block_w/2;
-    int sy= block_w*mb_y - block_w/2;
-    int x0= FFMAX(0,-sx);
-    int y0= FFMAX(0,-sy);
-    int x1= FFMIN(block_w*2, w-sx);
-    int y1= FFMIN(block_w*2, h-sy);
-    int i,x,y;
+    common_init(avctx);
+    alloc_blocks(s);
 
-    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);
+    s->version=0;
 
-    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;
-        }
+    s->m.avctx   = avctx;
+    s->m.flags   = avctx->flags;
+    s->m.bit_rate= avctx->bit_rate;
+
+    s->m.me.temp      =
+    s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
+    s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
+    s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
+    s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
+    h263_encode_init(&s->m); //mv_penalty
+
+    s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
+
+    if(avctx->flags&CODEC_FLAG_PASS1){
+        if(!avctx->stats_out)
+            avctx->stats_out = av_mallocz(256);
+    }
+    if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
+        if(ff_rate_control_init(&s->m) < 0)
+            return -1;
     }
+    s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
 
-    /* copy the regions where obmc[] = (uint8_t)256 */
-    if(LOG2_OBMC_MAX == 8
-        && (mb_x == 0 || mb_x == b_stride-1)
-        && (mb_y == 0 || mb_y == b_height-1)){
-        if(mb_x == 0)
-            x1 = block_w;
-        else
-            x0 = block_w;
-        if(mb_y == 0)
-            y1 = block_w;
-        else
-            y0 = block_w;
-        for(y=y0; y<y1; y++)
-            memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
+    avctx->coded_frame= &s->current_picture;
+    switch(avctx->pix_fmt){
+//    case PIX_FMT_YUV444P:
+//    case PIX_FMT_YUV422P:
+    case PIX_FMT_YUV420P:
+    case PIX_FMT_GRAY8:
+//    case PIX_FMT_YUV411P:
+//    case PIX_FMT_YUV410P:
+        s->colorspace_type= 0;
+        break;
+/*    case PIX_FMT_RGB32:
+        s->colorspace= 1;
+        break;*/
+    default:
+        av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
+        return -1;
     }
+//    avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
+    s->chroma_h_shift= 1;
+    s->chroma_v_shift= 1;
 
-    if(block_w==16){
-        /* FIXME rearrange dsputil to fit 32x32 cmp functions */
-        /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
-        /* FIXME cmps overlap but do not cover the wavelet's whole support.
-         * So improving the score of one block is not strictly guaranteed
-         * to improve the score of the whole frame, thus iterative motion
-         * estimation does not always converge. */
-        if(s->avctx->me_cmp == FF_CMP_W97)
-            distortion = w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
-        else if(s->avctx->me_cmp == FF_CMP_W53)
-            distortion = w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
-        else{
-            distortion = 0;
-            for(i=0; i<4; i++){
-                int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
-                distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
-            }
+    ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
+    ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
+
+    s->avctx->get_buffer(s->avctx, &s->input_picture);
+
+    if(s->avctx->me_method == ME_ITER){
+        int i;
+        int size= s->b_width * s->b_height << 2*s->block_max_depth;
+        for(i=0; i<s->max_ref_frames; i++){
+            s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
+            s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
         }
-    }else{
-        assert(block_w==8);
-        distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
     }
 
-    if(plane_index==0){
-        for(i=0; i<4; i++){
-/* ..RRr
- * .RXx.
- * rxx..
- */
-            rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
+    return 0;
+}
+
+//near copy & paste from dsputil, FIXME
+static int pix_sum(uint8_t * pix, int line_size, int w)
+{
+    int s, i, j;
+
+    s = 0;
+    for (i = 0; i < w; i++) {
+        for (j = 0; j < w; j++) {
+            s += pix[0];
+            pix ++;
         }
-        if(mb_x == b_stride-2)
-            rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
+        pix += line_size - w;
     }
-    return distortion + rate*penalty_factor;
+    return s;
 }
 
-static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
-    int i, y2;
-    Plane *p= &s->plane[plane_index];
-    const int block_size = MB_SIZE >> s->block_max_depth;
-    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 *dst= s->current_picture.data[plane_index];
-    uint8_t *src= s-> input_picture.data[plane_index];
-    //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
-    // const has only been removed from zero_dst to suppress a warning
-    static 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;
-    const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
+//near copy & paste from dsputil, FIXME
+static int pix_norm1(uint8_t * pix, int line_size, int w)
+{
+    int s, i, j;
+    uint32_t *sq = ff_squareTbl + 256;
 
-    for(i=0; i<9; i++){
-        int mb_x2= mb_x + (i%3) - 1;
-        int mb_y2= mb_y + (i/3) - 1;
-        int x= block_w*mb_x2 + block_w/2;
-        int y= block_w*mb_y2 + block_w/2;
-
-        add_yblock(s, 0, NULL, zero_dst, dst, obmc,
-                   x, y, block_w, block_w, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
-
-        //FIXME find a cleaner/simpler way to skip the outside stuff
-        for(y2= y; y2<0; y2++)
-            memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
-        for(y2= h; y2<y+block_w; y2++)
-            memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
-        if(x<0){
-            for(y2= y; y2<y+block_w; y2++)
-                memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
-        }
-        if(x+block_w > w){
-            for(y2= y; y2<y+block_w; y2++)
-                memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
-        }
-
-        assert(block_w== 8 || block_w==16);
-        distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_w);
-    }
-
-    if(plane_index==0){
-        BlockNode *b= &s->block[mb_x+mb_y*b_stride];
-        int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
-
-/* ..RRRr
- * .RXXx.
- * .RXXx.
- * rxxx.
- */
-        if(merged)
-            rate = get_block_bits(s, mb_x, mb_y, 2);
-        for(i=merged?4:0; i<9; i++){
-            static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
-            rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
+    s = 0;
+    for (i = 0; i < w; i++) {
+        for (j = 0; j < w; j ++) {
+            s += sq[pix[0]];
+            pix ++;
         }
+        pix += line_size - w;
     }
-    return distortion + rate*penalty_factor;
-}
-
-static av_always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3], int intra, const uint8_t *obmc_edged, int *best_rd){
-    const int b_stride= s->b_width << s->block_max_depth;
-    BlockNode *block= &s->block[mb_x + mb_y * b_stride];
-    BlockNode backup= *block;
-    int rd, index, value;
-
-    assert(mb_x>=0 && mb_y>=0);
-    assert(mb_x<b_stride);
-
-    if(intra){
-        block->color[0] = p[0];
-        block->color[1] = p[1];
-        block->color[2] = p[2];
-        block->type |= BLOCK_INTRA;
-    }else{
-        index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
-        value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
-        if(s->me_cache[index] == value)
-            return 0;
-        s->me_cache[index]= value;
-
-        block->mx= p[0];
-        block->my= p[1];
-        block->type &= ~BLOCK_INTRA;
-    }
-
-    rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
-
-//FIXME chroma
-    if(rd < *best_rd){
-        *best_rd= rd;
-        return 1;
-    }else{
-        *block= backup;
-        return 0;
-    }
-}
-
-/* special case for int[2] args we discard afterwards,
- * fixes compilation problem with gcc 2.95 */
-static av_always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, const uint8_t *obmc_edged, int *best_rd){
-    int p[2] = {p0, p1};
-    return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
+    return s;
 }
 
-static av_always_inline int check_4block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, int ref, int *best_rd){
-    const int b_stride= s->b_width << s->block_max_depth;
-    BlockNode *block= &s->block[mb_x + mb_y * b_stride];
-    BlockNode backup[4]= {block[0], block[1], block[b_stride], block[b_stride+1]};
-    int rd, index, value;
-
-    assert(mb_x>=0 && mb_y>=0);
-    assert(mb_x<b_stride);
-    assert(((mb_x|mb_y)&1) == 0);
-
-    index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
-    value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
-    if(s->me_cache[index] == value)
-        return 0;
-    s->me_cache[index]= value;
-
-    block->mx= p0;
-    block->my= p1;
-    block->ref= ref;
-    block->type &= ~BLOCK_INTRA;
-    block[1]= block[b_stride]= block[b_stride+1]= *block;
+//FIXME copy&paste
+#define P_LEFT P[1]
+#define P_TOP P[2]
+#define P_TOPRIGHT P[3]
+#define P_MEDIAN P[4]
+#define P_MV1 P[9]
+#define FLAG_QPEL   1 //must be 1
 
-    rd= get_4block_rd(s, mb_x, mb_y, 0);
+static int encode_q_branch(SnowContext *s, int level, int x, int y){
+    uint8_t p_buffer[1024];
+    uint8_t i_buffer[1024];
+    uint8_t p_state[sizeof(s->block_state)];
+    uint8_t i_state[sizeof(s->block_state)];
+    RangeCoder pc, ic;
+    uint8_t *pbbak= s->c.bytestream;
+    uint8_t *pbbak_start= s->c.bytestream_start;
+    int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
+    const int w= s->b_width  << s->block_max_depth;
+    const int h= s->b_height << s->block_max_depth;
+    const int rem_depth= s->block_max_depth - level;
+    const int index= (x + y*w) << rem_depth;
+    const int block_w= 1<<(LOG2_MB_SIZE - level);
+    int trx= (x+1)<<rem_depth;
+    int try= (y+1)<<rem_depth;
+    const BlockNode *left  = x ? &s->block[index-1] : &null_block;
+    const BlockNode *top   = y ? &s->block[index-w] : &null_block;
+    const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
+    const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
+    const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
+    const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
+    int pl = left->color[0];
+    int pcb= left->color[1];
+    int pcr= left->color[2];
+    int pmx, pmy;
+    int mx=0, my=0;
+    int l,cr,cb;
+    const int stride= s->current_picture.linesize[0];
+    const int uvstride= s->current_picture.linesize[1];
+    uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y*  stride)*block_w,
+                                s->input_picture.data[1] + (x + y*uvstride)*block_w/2,
+                                s->input_picture.data[2] + (x + y*uvstride)*block_w/2};
+    int P[10][2];
+    int16_t last_mv[3][2];
+    int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
+    const int shift= 1+qpel;
+    MotionEstContext *c= &s->m.me;
+    int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
+    int mx_context= av_log2(2*FFABS(left->mx - top->mx));
+    int my_context= av_log2(2*FFABS(left->my - top->my));
+    int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
+    int ref, best_ref, ref_score, ref_mx, ref_my;
 
-//FIXME chroma
-    if(rd < *best_rd){
-        *best_rd= rd;
-        return 1;
-    }else{
-        block[0]= backup[0];
-        block[1]= backup[1];
-        block[b_stride]= backup[2];
-        block[b_stride+1]= backup[3];
+    assert(sizeof(s->block_state) >= 256);
+    if(s->keyframe){
+        set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
         return 0;
     }
-}
-
-static void iterative_me(SnowContext *s){
-    int pass, mb_x, mb_y;
-    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;
-    int color[3];
-
-    {
-        RangeCoder r = s->c;
-        uint8_t state[sizeof(s->block_state)];
-        memcpy(state, s->block_state, sizeof(s->block_state));
-        for(mb_y= 0; mb_y<s->b_height; mb_y++)
-            for(mb_x= 0; mb_x<s->b_width; mb_x++)
-                encode_q_branch(s, 0, mb_x, mb_y);
-        s->c = r;
-        memcpy(s->block_state, state, sizeof(s->block_state));
-    }
-
-    for(pass=0; pass<25; pass++){
-        int change= 0;
-
-        for(mb_y= 0; mb_y<b_height; mb_y++){
-            for(mb_x= 0; mb_x<b_width; mb_x++){
-                int dia_change, i, j, ref;
-                int best_rd= INT_MAX, ref_rd;
-                BlockNode backup, ref_b;
-                const int index= mb_x + mb_y * b_stride;
-                BlockNode *block= &s->block[index];
-                BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
-                BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
-                BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
-                BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
-                BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
-                BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
-                BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
-                BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
-                const int b_w= (MB_SIZE >> s->block_max_depth);
-                uint8_t obmc_edged[b_w*2][b_w*2];
-
-                if(pass && (block->type & BLOCK_OPT))
-                    continue;
-                block->type |= BLOCK_OPT;
-
-                backup= *block;
-
-                if(!s->me_cache_generation)
-                    memset(s->me_cache, 0, sizeof(s->me_cache));
-                s->me_cache_generation += 1<<22;
-
-                //FIXME precalculate
-                {
-                    int x, y;
-                    memcpy(obmc_edged, obmc_tab[s->block_max_depth], b_w*b_w*4);
-                    if(mb_x==0)
-                        for(y=0; y<b_w*2; y++)
-                            memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
-                    if(mb_x==b_stride-1)
-                        for(y=0; y<b_w*2; y++)
-                            memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
-                    if(mb_y==0){
-                        for(x=0; x<b_w*2; x++)
-                            obmc_edged[0][x] += obmc_edged[b_w-1][x];
-                        for(y=1; y<b_w; y++)
-                            memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
-                    }
-                    if(mb_y==b_height-1){
-                        for(x=0; x<b_w*2; x++)
-                            obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
-                        for(y=b_w; y<b_w*2-1; y++)
-                            memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
-                    }
-                }
 
-                //skip stuff outside the picture
-                if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
-                    uint8_t *src= s->  input_picture.data[0];
-                    uint8_t *dst= s->current_picture.data[0];
-                    const int stride= s->current_picture.linesize[0];
-                    const int block_w= MB_SIZE >> s->block_max_depth;
-                    const int sx= block_w*mb_x - block_w/2;
-                    const int sy= block_w*mb_y - block_w/2;
-                    const int w= s->plane[0].width;
-                    const int h= s->plane[0].height;
-                    int y;
-
-                    for(y=sy; y<0; y++)
-                        memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
-                    for(y=h; y<sy+block_w*2; y++)
-                        memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
-                    if(sx<0){
-                        for(y=sy; y<sy+block_w*2; y++)
-                            memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
-                    }
-                    if(sx+block_w*2 > w){
-                        for(y=sy; y<sy+block_w*2; y++)
-                            memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
-                    }
-                }
-
-                // intra(black) = neighbors' contribution to the current block
-                for(i=0; i<3; i++)
-                    color[i]= get_dc(s, mb_x, mb_y, i);
-
-                // get previous score (cannot be cached due to OBMC)
-                if(pass > 0 && (block->type&BLOCK_INTRA)){
-                    int color0[3]= {block->color[0], block->color[1], block->color[2]};
-                    check_block(s, mb_x, mb_y, color0, 1, *obmc_edged, &best_rd);
-                }else
-                    check_block_inter(s, mb_x, mb_y, block->mx, block->my, *obmc_edged, &best_rd);
+//    clip predictors / edge ?
 
-                ref_b= *block;
-                ref_rd= best_rd;
-                for(ref=0; ref < s->ref_frames; ref++){
-                    int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
-                    if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
-                        continue;
-                    block->ref= ref;
-                    best_rd= INT_MAX;
+    P_LEFT[0]= left->mx;
+    P_LEFT[1]= left->my;
+    P_TOP [0]= top->mx;
+    P_TOP [1]= top->my;
+    P_TOPRIGHT[0]= tr->mx;
+    P_TOPRIGHT[1]= tr->my;
 
-                    check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], *obmc_edged, &best_rd);
-                    check_block_inter(s, mb_x, mb_y, 0, 0, *obmc_edged, &best_rd);
-                    if(tb)
-                        check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], *obmc_edged, &best_rd);
-                    if(lb)
-                        check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], *obmc_edged, &best_rd);
-                    if(rb)
-                        check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], *obmc_edged, &best_rd);
-                    if(bb)
-                        check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], *obmc_edged, &best_rd);
+    last_mv[0][0]= s->block[index].mx;
+    last_mv[0][1]= s->block[index].my;
+    last_mv[1][0]= right->mx;
+    last_mv[1][1]= right->my;
+    last_mv[2][0]= bottom->mx;
+    last_mv[2][1]= bottom->my;
 
-                    /* fullpel ME */
-                    //FIXME avoid subpel interpolation / round to nearest integer
-                    do{
-                        dia_change=0;
-                        for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
-                            for(j=0; j<i; j++){
-                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), *obmc_edged, &best_rd);
-                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), *obmc_edged, &best_rd);
-                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), *obmc_edged, &best_rd);
-                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), *obmc_edged, &best_rd);
-                            }
-                        }
-                    }while(dia_change);
-                    /* subpel ME */
-                    do{
-                        static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
-                        dia_change=0;
-                        for(i=0; i<8; i++)
-                            dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], *obmc_edged, &best_rd);
-                    }while(dia_change);
-                    //FIXME or try the standard 2 pass qpel or similar
+    s->m.mb_stride=2;
+    s->m.mb_x=
+    s->m.mb_y= 0;
+    c->skip= 0;
 
-                    mvr[0][0]= block->mx;
-                    mvr[0][1]= block->my;
-                    if(ref_rd > best_rd){
-                        ref_rd= best_rd;
-                        ref_b= *block;
-                    }
-                }
-                best_rd= ref_rd;
-                *block= ref_b;
-#if 1
-                check_block(s, mb_x, mb_y, color, 1, *obmc_edged, &best_rd);
-                //FIXME RD style color selection
-#endif
-                if(!same_block(block, &backup)){
-                    if(tb ) tb ->type &= ~BLOCK_OPT;
-                    if(lb ) lb ->type &= ~BLOCK_OPT;
-                    if(rb ) rb ->type &= ~BLOCK_OPT;
-                    if(bb ) bb ->type &= ~BLOCK_OPT;
-                    if(tlb) tlb->type &= ~BLOCK_OPT;
-                    if(trb) trb->type &= ~BLOCK_OPT;
-                    if(blb) blb->type &= ~BLOCK_OPT;
-                    if(brb) brb->type &= ~BLOCK_OPT;
-                    change ++;
-                }
-            }
-        }
-        av_log(NULL, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
-        if(!change)
-            break;
-    }
+    assert(c->  stride ==   stride);
+    assert(c->uvstride == uvstride);
 
-    if(s->block_max_depth == 1){
-        int change= 0;
-        for(mb_y= 0; mb_y<b_height; mb_y+=2){
-            for(mb_x= 0; mb_x<b_width; mb_x+=2){
-                int i;
-                int best_rd, init_rd;
-                const int index= mb_x + mb_y * b_stride;
-                BlockNode *b[4];
+    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
+    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
+    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
+    c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
 
-                b[0]= &s->block[index];
-                b[1]= b[0]+1;
-                b[2]= b[0]+b_stride;
-                b[3]= b[2]+1;
-                if(same_block(b[0], b[1]) &&
-                   same_block(b[0], b[2]) &&
-                   same_block(b[0], b[3]))
-                    continue;
+    c->xmin = - x*block_w - 16+3;
+    c->ymin = - y*block_w - 16+3;
+    c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
+    c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 
-                if(!s->me_cache_generation)
-                    memset(s->me_cache, 0, sizeof(s->me_cache));
-                s->me_cache_generation += 1<<22;
+    if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
+    if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
+    if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
+    if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
+    if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
+    if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
+    if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 
-                init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
+    P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
+    P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 
-                //FIXME more multiref search?
-                check_4block_inter(s, mb_x, mb_y,
-                                   (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
-                                   (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
+    if (!y) {
+        c->pred_x= P_LEFT[0];
+        c->pred_y= P_LEFT[1];
+    } else {
+        c->pred_x = P_MEDIAN[0];
+        c->pred_y = P_MEDIAN[1];
+    }
 
-                for(i=0; i<4; i++)
-                    if(!(b[i]->type&BLOCK_INTRA))
-                        check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
+    score= INT_MAX;
+    best_ref= 0;
+    for(ref=0; ref<s->ref_frames; ref++){
+        init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
 
-                if(init_rd != best_rd)
-                    change++;
-            }
+        ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
+                                         (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
+
+        assert(ref_mx >= c->xmin);
+        assert(ref_mx <= c->xmax);
+        assert(ref_my >= c->ymin);
+        assert(ref_my <= c->ymax);
+
+        ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
+        ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
+        ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
+        if(s->ref_mvs[ref]){
+            s->ref_mvs[ref][index][0]= ref_mx;
+            s->ref_mvs[ref][index][1]= ref_my;
+            s->ref_scores[ref][index]= ref_score;
+        }
+        if(score > ref_score){
+            score= ref_score;
+            best_ref= ref;
+            mx= ref_mx;
+            my= ref_my;
         }
-        av_log(NULL, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
     }
-}
+    //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 
-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;
+  //  subpel search
+    base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
+    pc= s->c;
+    pc.bytestream_start=
+    pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
+    memcpy(p_state, s->block_state, sizeof(s->block_state));
 
-    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;
+    if(level!=s->block_max_depth)
+        put_rac(&pc, &p_state[4 + s_context], 1);
+    put_rac(&pc, &p_state[1 + left->type + top->type], 0);
+    if(s->ref_frames > 1)
+        put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
+    pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
+    put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
+    put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
+    p_len= pc.bytestream - pc.bytestream_start;
+    score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
+
+    block_s= block_w*block_w;
+    sum = pix_sum(current_data[0], stride, block_w);
+    l= (sum + block_s/2)/block_s;
+    iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
+
+    block_s= block_w*block_w>>2;
+    sum = pix_sum(current_data[1], uvstride, block_w>>1);
+    cb= (sum + block_s/2)/block_s;
+//    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
+    sum = pix_sum(current_data[2], uvstride, block_w>>1);
+    cr= (sum + block_s/2)/block_s;
+//    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
+
+    ic= s->c;
+    ic.bytestream_start=
+    ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
+    memcpy(i_state, s->block_state, sizeof(s->block_state));
+    if(level!=s->block_max_depth)
+        put_rac(&ic, &i_state[4 + s_context], 1);
+    put_rac(&ic, &i_state[1 + left->type + top->type], 1);
+    put_symbol(&ic, &i_state[32],  l-pl , 1);
+    put_symbol(&ic, &i_state[64], cb-pcb, 1);
+    put_symbol(&ic, &i_state[96], cr-pcr, 1);
+    i_len= ic.bytestream - ic.bytestream_start;
+    iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
+
+//    assert(score==256*256*256*64-1);
+    assert(iscore < 255*255*256 + s->lambda2*10);
+    assert(iscore >= 0);
+    assert(l>=0 && l<=255);
+    assert(pl>=0 && pl<=255);
+
+    if(level==0){
+        int varc= iscore >> 8;
+        int vard= score >> 8;
+        if (vard <= 64 || vard < varc)
+            c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
+        else
+            c->scene_change_score+= s->m.qscale;
     }
 
-    bias= bias ? 0 : (3*qmul)>>3;
-    thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
-    thres2= 2*thres1;
+    if(level!=s->block_max_depth){
+        put_rac(&s->c, &s->block_state[4 + s_context], 0);
+        score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
+        score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
+        score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
+        score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
+        score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 
-    if(!bias){
-        for(y=0; y<h; y++){
-            for(x=0; x<w; x++){
-                int i= src[x + y*stride];
+        if(score2 < score && score2 < iscore)
+            return score2;
+    }
 
-                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;
-            }
-        }
+    if(iscore < score){
+        pred_mv(s, &pmx, &pmy, 0, left, top, tr);
+        memcpy(pbbak, i_buffer, i_len);
+        s->c= ic;
+        s->c.bytestream_start= pbbak_start;
+        s->c.bytestream= pbbak + i_len;
+        set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
+        memcpy(s->block_state, i_state, sizeof(s->block_state));
+        return iscore;
     }else{
-        for(y=0; y<h; y++){
-            for(x=0; x<w; x++){
-                int i= src[x + y*stride];
-
-                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;
-            }
-        }
+        memcpy(pbbak, p_buffer, p_len);
+        s->c= pc;
+        s->c.bytestream_start= pbbak_start;
+        s->c.bytestream= pbbak + p_len;
+        set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
+        memcpy(s->block_state, p_state, sizeof(s->block_state));
+        return score;
     }
 }
 
-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;
+static void encode_q_branch2(SnowContext *s, int level, int x, int y){
+    const int w= s->b_width  << s->block_max_depth;
+    const int rem_depth= s->block_max_depth - level;
+    const int index= (x + y*w) << rem_depth;
+    int trx= (x+1)<<rem_depth;
+    BlockNode *b= &s->block[index];
+    const BlockNode *left  = x ? &s->block[index-1] : &null_block;
+    const BlockNode *top   = y ? &s->block[index-w] : &null_block;
+    const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
+    const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
+    int pl = left->color[0];
+    int pcb= left->color[1];
+    int pcr= left->color[2];
+    int pmx, pmy;
+    int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
+    int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
+    int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
+    int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 
-    if(s->qlog == LOSSLESS_QLOG) return;
+    if(s->keyframe){
+        set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
+        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){
-                line[x]=  (( i*qmul + qadd)>>(QEXPSHIFT));
-            }
+    if(level!=s->block_max_depth){
+        if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
+            put_rac(&s->c, &s->block_state[4 + s_context], 1);
+        }else{
+            put_rac(&s->c, &s->block_state[4 + s_context], 0);
+            encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
+            encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
+            encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
+            encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
+            return;
         }
     }
+    if(b->type & BLOCK_INTRA){
+        pred_mv(s, &pmx, &pmy, 0, left, top, tr);
+        put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
+        put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
+        put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
+        put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
+        set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
+    }else{
+        pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
+        put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
+        if(s->ref_frames > 1)
+            put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
+        put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
+        put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
+        set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
+    }
 }
 
-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;
+static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
+    int i, x2, y2;
+    Plane *p= &s->plane[plane_index];
+    const int block_size = MB_SIZE >> s->block_max_depth;
+    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];
+    BlockNode backup= *b;
+    int ab=0;
+    int aa=0;
 
-    if(s->qlog == LOSSLESS_QLOG) return;
+    b->type|= BLOCK_INTRA;
+    b->color[plane_index]= 0;
+    memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 
-    for(y=0; y<h; y++){
-        for(x=0; x<w; x++){
-            int i= src[x + y*stride];
-            if(i<0){
-                src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
-            }else if(i>0){
-                src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
-            }
-        }
-    }
-}
+    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;
+        int y= block_w*mb_y2 + block_w/2;
 
-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;
+        add_yblock(s, 0, NULL, dst + ((i&1)+(i>>1)*obmc_stride)*block_w, NULL, obmc,
+                    x, y, block_w, block_w, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 
-    for(y=h-1; y>=0; y--){
-        for(x=w-1; x>=0; x--){
-            int i= x + y*stride;
+        for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_w); y2++){
+            for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
+                int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_w*mb_y - block_w/2))*obmc_stride;
+                int obmc_v= obmc[index];
+                int d;
+                if(y<0) obmc_v += obmc[index + block_w*obmc_stride];
+                if(x<0) obmc_v += obmc[index + block_w];
+                if(y+block_w>h) obmc_v += obmc[index - block_w*obmc_stride];
+                if(x+block_w>w) obmc_v += obmc[index - block_w];
+                //FIXME precalculate this or simplify it somehow else
 
-            if(x){
-                if(use_median){
-                    if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
-                    else  src[i] -= src[i - 1];
-                }else{
-                    if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
-                    else  src[i] -= src[i - 1];
-                }
-            }else{
-                if(y) src[i] -= src[i - stride];
+                d = -dst[index] + (1<<(FRAC_BITS-1));
+                dst[index] = d;
+                ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
+                aa += obmc_v * obmc_v; //FIXME precalculate this
             }
         }
     }
-}
-
-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;
+    *b= backup;
 
-    IDWTELEM * line=0; // silence silly "could be used without having been initialized" warning
-    IDWTELEM * prev;
+    return av_clip(((ab<<LOG2_OBMC_MAX) + aa/2)/aa, 0, 255); //FIXME we should not need clipping
+}
 
-    if (start_y != 0)
-        line = slice_buffer_get_line(sb, ((start_y - 1) * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;
+static inline int get_block_bits(SnowContext *s, int x, int y, int w){
+    const int b_stride = s->b_width << s->block_max_depth;
+    const int b_height = s->b_height<< s->block_max_depth;
+    int index= x + y*b_stride;
+    const BlockNode *b     = &s->block[index];
+    const BlockNode *left  = x ? &s->block[index-1] : &null_block;
+    const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
+    const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
+    const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
+    int dmx, dmy;
+//  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
+//  int my_context= av_log2(2*FFABS(left->my - top->my));
 
-    for(y=start_y; y<end_y; y++){
-        prev = line;
-//        line = slice_buffer_get_line_from_address(sb, src + (y * stride));
-        line = slice_buffer_get_line(sb, (y * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;
-        for(x=0; x<w; x++){
-            if(x){
-                if(use_median){
-                    if(y && x+1<w) line[x] += mid_pred(line[x - 1], prev[x], prev[x + 1]);
-                    else  line[x] += line[x - 1];
-                }else{
-                    if(y) line[x] += mid_pred(line[x - 1], prev[x], line[x - 1] + prev[x] - prev[x - 1]);
-                    else  line[x] += line[x - 1];
-                }
-            }else{
-                if(y) line[x] += prev[x];
-            }
-        }
+    if(x<0 || x>=b_stride || y>=b_height)
+        return 0;
+/*
+1            0      0
+01X          1-2    1
+001XX        3-6    2-3
+0001XXX      7-14   4-7
+00001XXXX   15-30   8-15
+*/
+//FIXME try accurate rate
+//FIXME intra and inter predictors if surrounding blocks are not the same type
+    if(b->type & BLOCK_INTRA){
+        return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
+                   + av_log2(2*FFABS(left->color[1] - b->color[1]))
+                   + av_log2(2*FFABS(left->color[2] - b->color[2])));
+    }else{
+        pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
+        dmx-= b->mx;
+        dmy-= b->my;
+        return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
+                    + av_log2(2*FFABS(dmy))
+                    + av_log2(2*b->ref));
     }
 }
 
-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;
+static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, const uint8_t *obmc_edged){
+    Plane *p= &s->plane[plane_index];
+    const int block_size = MB_SIZE >> s->block_max_depth;
+    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 = s->scratchbuf;
+    uint8_t tmp[ref_stride*(2*MB_SIZE+HTAPS_MAX-1)];
+    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;
+    const int h= p->height;
+    int distortion;
+    int rate= 0;
+    const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
+    int sx= block_w*mb_x - block_w/2;
+    int sy= block_w*mb_y - block_w/2;
+    int x0= FFMAX(0,-sx);
+    int y0= FFMAX(0,-sy);
+    int x1= FFMIN(block_w*2, w-sx);
+    int y1= FFMIN(block_w*2, h-sy);
+    int i,x,y;
 
-    for(y=0; y<h; y++){
-        for(x=0; x<w; x++){
-            int i= x + y*stride;
+    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);
 
-            if(x){
-                if(use_median){
-                    if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
-                    else  src[i] += src[i - 1];
-                }else{
-                    if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
-                    else  src[i] += src[i - 1];
-                }
-            }else{
-                if(y) src[i] += src[i - stride];
-            }
+    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;
         }
     }
-}
 
-static void encode_qlogs(SnowContext *s){
-    int plane_index, level, orientation;
+    /* copy the regions where obmc[] = (uint8_t)256 */
+    if(LOG2_OBMC_MAX == 8
+        && (mb_x == 0 || mb_x == b_stride-1)
+        && (mb_y == 0 || mb_y == b_height-1)){
+        if(mb_x == 0)
+            x1 = block_w;
+        else
+            x0 = block_w;
+        if(mb_y == 0)
+            y1 = block_w;
+        else
+            y0 = block_w;
+        for(y=y0; y<y1; y++)
+            memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
+    }
 
-    for(plane_index=0; plane_index<2; plane_index++){
-        for(level=0; level<s->spatial_decomposition_count; level++){
-            for(orientation=level ? 1:0; orientation<4; orientation++){
-                if(orientation==2) continue;
-                put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
+    if(block_w==16){
+        /* FIXME rearrange dsputil to fit 32x32 cmp functions */
+        /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
+        /* FIXME cmps overlap but do not cover the wavelet's whole support.
+         * So improving the score of one block is not strictly guaranteed
+         * to improve the score of the whole frame, thus iterative motion
+         * estimation does not always converge. */
+        if(s->avctx->me_cmp == FF_CMP_W97)
+            distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
+        else if(s->avctx->me_cmp == FF_CMP_W53)
+            distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
+        else{
+            distortion = 0;
+            for(i=0; i<4; i++){
+                int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
+                distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
             }
         }
+    }else{
+        assert(block_w==8);
+        distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
     }
-}
-
-static void encode_header(SnowContext *s){
-    int plane_index, i;
-    uint8_t kstate[32];
-
-    memset(kstate, MID_STATE, sizeof(kstate));
 
-    put_rac(&s->c, kstate, s->keyframe);
-    if(s->keyframe || s->always_reset){
-        reset_contexts(s);
-        s->last_spatial_decomposition_type=
-        s->last_qlog=
-        s->last_qbias=
-        s->last_mv_scale=
-        s->last_block_max_depth= 0;
-        for(plane_index=0; plane_index<2; plane_index++){
-            Plane *p= &s->plane[plane_index];
-            p->last_htaps=0;
-            p->last_diag_mc=0;
-            memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
+    if(plane_index==0){
+        for(i=0; i<4; i++){
+/* ..RRr
+ * .RXx.
+ * rxx..
+ */
+            rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
         }
+        if(mb_x == b_stride-2)
+            rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
     }
-    if(s->keyframe){
-        put_symbol(&s->c, s->header_state, s->version, 0);
-        put_rac(&s->c, s->header_state, s->always_reset);
-        put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
-        put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
-        put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
-        put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
-        put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
-        put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
-        put_rac(&s->c, s->header_state, s->spatial_scalability);
-//        put_rac(&s->c, s->header_state, s->rate_scalability);
-        put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
+    return distortion + rate*penalty_factor;
+}
 
-        encode_qlogs(s);
-    }
+static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
+    int i, y2;
+    Plane *p= &s->plane[plane_index];
+    const int block_size = MB_SIZE >> s->block_max_depth;
+    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 *dst= s->current_picture.data[plane_index];
+    uint8_t *src= s-> input_picture.data[plane_index];
+    //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
+    // const has only been removed from zero_dst to suppress a warning
+    static 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;
+    const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 
-    if(!s->keyframe){
-        int update_mc=0;
-        for(plane_index=0; plane_index<2; plane_index++){
-            Plane *p= &s->plane[plane_index];
-            update_mc |= p->last_htaps   != p->htaps;
-            update_mc |= p->last_diag_mc != p->diag_mc;
-            update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
+    for(i=0; i<9; i++){
+        int mb_x2= mb_x + (i%3) - 1;
+        int mb_y2= mb_y + (i/3) - 1;
+        int x= block_w*mb_x2 + block_w/2;
+        int y= block_w*mb_y2 + block_w/2;
+
+        add_yblock(s, 0, NULL, zero_dst, dst, obmc,
+                   x, y, block_w, block_w, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
+
+        //FIXME find a cleaner/simpler way to skip the outside stuff
+        for(y2= y; y2<0; y2++)
+            memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
+        for(y2= h; y2<y+block_w; y2++)
+            memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
+        if(x<0){
+            for(y2= y; y2<y+block_w; y2++)
+                memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
         }
-        put_rac(&s->c, s->header_state, update_mc);
-        if(update_mc){
-            for(plane_index=0; plane_index<2; plane_index++){
-                Plane *p= &s->plane[plane_index];
-                put_rac(&s->c, s->header_state, p->diag_mc);
-                put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
-                for(i= p->htaps/2; i; i--)
-                    put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
-            }
+        if(x+block_w > w){
+            for(y2= y; y2<y+block_w; y2++)
+                memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
         }
-        if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
-            put_rac(&s->c, s->header_state, 1);
-            put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
-            encode_qlogs(s);
-        }else
-            put_rac(&s->c, s->header_state, 0);
-    }
-
-    put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
-    put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
-    put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
-    put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
-    put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
 
-}
+        assert(block_w== 8 || block_w==16);
+        distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_w);
+    }
 
-static void update_last_header_values(SnowContext *s){
-    int plane_index;
+    if(plane_index==0){
+        BlockNode *b= &s->block[mb_x+mb_y*b_stride];
+        int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 
-    if(!s->keyframe){
-        for(plane_index=0; plane_index<2; plane_index++){
-            Plane *p= &s->plane[plane_index];
-            p->last_diag_mc= p->diag_mc;
-            p->last_htaps  = p->htaps;
-            memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
+/* ..RRRr
+ * .RXXx.
+ * .RXXx.
+ * rxxx.
+ */
+        if(merged)
+            rate = get_block_bits(s, mb_x, mb_y, 2);
+        for(i=merged?4:0; i<9; i++){
+            static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
+            rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
         }
     }
-
-    s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
-    s->last_qlog                        = s->qlog;
-    s->last_qbias                       = s->qbias;
-    s->last_mv_scale                    = s->mv_scale;
-    s->last_block_max_depth             = s->block_max_depth;
-    s->last_spatial_decomposition_count = s->spatial_decomposition_count;
+    return distortion + rate*penalty_factor;
 }
 
-static void decode_qlogs(SnowContext *s){
-    int plane_index, level, orientation;
+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;
 
-    for(plane_index=0; plane_index<3; plane_index++){
-        for(level=0; level<s->spatial_decomposition_count; level++){
-            for(orientation=level ? 1:0; orientation<4; orientation++){
-                int q;
-                if     (plane_index==2) q= s->plane[1].band[level][orientation].qlog;
-                else if(orientation==2) q= s->plane[plane_index].band[level][1].qlog;
-                else                    q= get_symbol(&s->c, s->header_state, 1);
-                s->plane[plane_index].band[level][orientation].qlog= q;
+    if(1){
+        int run=0;
+        int runs[w*h];
+        int run_index=0;
+        int max_index;
+
+        for(y=0; y<h; y++){
+            for(x=0; x<w; x++){
+                int v, p=0;
+                int /*ll=0, */l=0, lt=0, t=0, rt=0;
+                v= src[x + y*stride];
+
+                if(y){
+                    t= src[x + (y-1)*stride];
+                    if(x){
+                        lt= src[x - 1 + (y-1)*stride];
+                    }
+                    if(x + 1 < w){
+                        rt= src[x + 1 + (y-1)*stride];
+                    }
+                }
+                if(x){
+                    l= src[x - 1 + y*stride];
+                    /*if(x > 1){
+                        if(orientation==1) ll= src[y + (x-2)*stride];
+                        else               ll= src[x - 2 + y*stride];
+                    }*/
+                }
+                if(parent){
+                    int px= x>>1;
+                    int py= y>>1;
+                    if(px<b->parent->width && py<b->parent->height)
+                        p= parent[px + py*2*stride];
+                }
+                if(!(/*ll|*/l|lt|t|rt|p)){
+                    if(v){
+                        runs[run_index++]= run;
+                        run=0;
+                    }else{
+                        run++;
+                    }
+                }
             }
         }
-    }
-}
+        max_index= run_index;
+        runs[run_index++]= run;
+        run_index=0;
+        run= runs[run_index++];
 
-#define GET_S(dst, check) \
-    tmp= get_symbol(&s->c, s->header_state, 0);\
-    if(!(check)){\
-        av_log(s->avctx, AV_LOG_ERROR, "Error " #dst " is %d\n", tmp);\
-        return -1;\
-    }\
-    dst= tmp;
+        put_symbol2(&s->c, b->state[30], max_index, 0);
+        if(run_index <= max_index)
+            put_symbol2(&s->c, b->state[1], run, 3);
 
-static int decode_header(SnowContext *s){
-    int plane_index, tmp;
-    uint8_t kstate[32];
+        for(y=0; y<h; y++){
+            if(s->c.bytestream_end - s->c.bytestream < w*40){
+                av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
+                return -1;
+            }
+            for(x=0; x<w; x++){
+                int v, p=0;
+                int /*ll=0, */l=0, lt=0, t=0, rt=0;
+                v= src[x + y*stride];
 
-    memset(kstate, MID_STATE, sizeof(kstate));
+                if(y){
+                    t= src[x + (y-1)*stride];
+                    if(x){
+                        lt= src[x - 1 + (y-1)*stride];
+                    }
+                    if(x + 1 < w){
+                        rt= src[x + 1 + (y-1)*stride];
+                    }
+                }
+                if(x){
+                    l= src[x - 1 + y*stride];
+                    /*if(x > 1){
+                        if(orientation==1) ll= src[y + (x-2)*stride];
+                        else               ll= src[x - 2 + y*stride];
+                    }*/
+                }
+                if(parent){
+                    int px= x>>1;
+                    int py= y>>1;
+                    if(px<b->parent->width && py<b->parent->height)
+                        p= parent[px + py*2*stride];
+                }
+                if(/*ll|*/l|lt|t|rt|p){
+                    int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 
-    s->keyframe= get_rac(&s->c, kstate);
-    if(s->keyframe || s->always_reset){
-        reset_contexts(s);
-        s->spatial_decomposition_type=
-        s->qlog=
-        s->qbias=
-        s->mv_scale=
-        s->block_max_depth= 0;
-    }
-    if(s->keyframe){
-        GET_S(s->version, tmp <= 0U)
-        s->always_reset= get_rac(&s->c, s->header_state);
-        s->temporal_decomposition_type= get_symbol(&s->c, s->header_state, 0);
-        s->temporal_decomposition_count= get_symbol(&s->c, s->header_state, 0);
-        s->spatial_decomposition_count= get_symbol(&s->c, s->header_state, 0);
-        s->colorspace_type= get_symbol(&s->c, s->header_state, 0);
-        s->chroma_h_shift= get_symbol(&s->c, s->header_state, 0);
-        s->chroma_v_shift= get_symbol(&s->c, s->header_state, 0);
-        s->spatial_scalability= get_rac(&s->c, s->header_state);
-//        s->rate_scalability= get_rac(&s->c, s->header_state);
-        GET_S(s->max_ref_frames, tmp < (unsigned)MAX_REF_FRAMES)
-        s->max_ref_frames++;
+                    put_rac(&s->c, &b->state[0][context], !!v);
+                }else{
+                    if(!run){
+                        run= runs[run_index++];
 
-        decode_qlogs(s);
-    }
+                        if(run_index <= max_index)
+                            put_symbol2(&s->c, b->state[1], run, 3);
+                        assert(v);
+                    }else{
+                        run--;
+                        assert(!v);
+                    }
+                }
+                if(v){
+                    int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
+                    int l2= 2*FFABS(l) + (l<0);
+                    int t2= 2*FFABS(t) + (t<0);
 
-    if(!s->keyframe){
-        if(get_rac(&s->c, s->header_state)){
-            for(plane_index=0; plane_index<2; plane_index++){
-                int htaps, i, sum=0;
-                Plane *p= &s->plane[plane_index];
-                p->diag_mc= get_rac(&s->c, s->header_state);
-                htaps= get_symbol(&s->c, s->header_state, 0)*2 + 2;
-                if((unsigned)htaps > HTAPS_MAX || htaps==0)
-                    return -1;
-                p->htaps= htaps;
-                for(i= htaps/2; i; i--){
-                    p->hcoeff[i]= get_symbol(&s->c, s->header_state, 0) * (1-2*(i&1));
-                    sum += p->hcoeff[i];
+                    put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
+                    put_rac(&s->c, &b->state[0][16 + 1 + 3 + quant3bA[l2&0xFF] + 3*quant3bA[t2&0xFF]], v<0);
                 }
-                p->hcoeff[0]= 32-sum;
             }
-            s->plane[2].diag_mc= s->plane[1].diag_mc;
-            s->plane[2].htaps  = s->plane[1].htaps;
-            memcpy(s->plane[2].hcoeff, s->plane[1].hcoeff, sizeof(s->plane[1].hcoeff));
-        }
-        if(get_rac(&s->c, s->header_state)){
-            s->spatial_decomposition_count= get_symbol(&s->c, s->header_state, 0);
-            decode_qlogs(s);
         }
     }
-
-    s->spatial_decomposition_type+= get_symbol(&s->c, s->header_state, 1);
-    if(s->spatial_decomposition_type > 1U){
-        av_log(s->avctx, AV_LOG_ERROR, "spatial_decomposition_type %d not supported", s->spatial_decomposition_type);
-        return -1;
-    }
-
-    s->qlog           += get_symbol(&s->c, s->header_state, 1);
-    s->mv_scale       += get_symbol(&s->c, s->header_state, 1);
-    s->qbias          += get_symbol(&s->c, s->header_state, 1);
-    s->block_max_depth+= get_symbol(&s->c, s->header_state, 1);
-    if(s->block_max_depth > 1 || s->block_max_depth < 0){
-        av_log(s->avctx, AV_LOG_ERROR, "block_max_depth= %d is too large", s->block_max_depth);
-        s->block_max_depth= 0;
-        return -1;
-    }
-
     return 0;
 }
 
-static void init_qexp(void){
-    int i;
-    double v=128;
-
-    for(i=0; i<QROOT; i++){
-        qexp[i]= lrintf(v);
-        v *= pow(2, 1.0 / QROOT);
-    }
+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);
 }
 
-static av_cold int common_init(AVCodecContext *avctx){
-    SnowContext *s = avctx->priv_data;
-    int width, height;
-    int i, j;
-
-    s->avctx= avctx;
-    s->max_ref_frames=1; //just make sure its not an invalid value in case of no initial keyframe
+static av_always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3], int intra, const uint8_t *obmc_edged, int *best_rd){
+    const int b_stride= s->b_width << s->block_max_depth;
+    BlockNode *block= &s->block[mb_x + mb_y * b_stride];
+    BlockNode backup= *block;
+    int rd, index, value;
 
-    dsputil_init(&s->dsp, avctx);
+    assert(mb_x>=0 && mb_y>=0);
+    assert(mb_x<b_stride);
 
-#define mcf(dx,dy)\
-    s->dsp.put_qpel_pixels_tab       [0][dy+dx/4]=\
-    s->dsp.put_no_rnd_qpel_pixels_tab[0][dy+dx/4]=\
-        s->dsp.put_h264_qpel_pixels_tab[0][dy+dx/4];\
-    s->dsp.put_qpel_pixels_tab       [1][dy+dx/4]=\
-    s->dsp.put_no_rnd_qpel_pixels_tab[1][dy+dx/4]=\
-        s->dsp.put_h264_qpel_pixels_tab[1][dy+dx/4];
+    if(intra){
+        block->color[0] = p[0];
+        block->color[1] = p[1];
+        block->color[2] = p[2];
+        block->type |= BLOCK_INTRA;
+    }else{
+        index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
+        value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
+        if(s->me_cache[index] == value)
+            return 0;
+        s->me_cache[index]= value;
 
-    mcf( 0, 0)
-    mcf( 4, 0)
-    mcf( 8, 0)
-    mcf(12, 0)
-    mcf( 0, 4)
-    mcf( 4, 4)
-    mcf( 8, 4)
-    mcf(12, 4)
-    mcf( 0, 8)
-    mcf( 4, 8)
-    mcf( 8, 8)
-    mcf(12, 8)
-    mcf( 0,12)
-    mcf( 4,12)
-    mcf( 8,12)
-    mcf(12,12)
+        block->mx= p[0];
+        block->my= p[1];
+        block->type &= ~BLOCK_INTRA;
+    }
 
-#define mcfh(dx,dy)\
-    s->dsp.put_pixels_tab       [0][dy/4+dx/8]=\
-    s->dsp.put_no_rnd_pixels_tab[0][dy/4+dx/8]=\
-        mc_block_hpel ## dx ## dy ## 16;\
-    s->dsp.put_pixels_tab       [1][dy/4+dx/8]=\
-    s->dsp.put_no_rnd_pixels_tab[1][dy/4+dx/8]=\
-        mc_block_hpel ## dx ## dy ## 8;
+    rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
 
-    mcfh(0, 0)
-    mcfh(8, 0)
-    mcfh(0, 8)
-    mcfh(8, 8)
+//FIXME chroma
+    if(rd < *best_rd){
+        *best_rd= rd;
+        return 1;
+    }else{
+        *block= backup;
+        return 0;
+    }
+}
 
-    if(!qexp[0])
-        init_qexp();
+/* special case for int[2] args we discard afterwards,
+ * fixes compilation problem with gcc 2.95 */
+static av_always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, const uint8_t *obmc_edged, int *best_rd){
+    int p[2] = {p0, p1};
+    return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
+}
 
-//    dec += FFMAX(s->chroma_h_shift, s->chroma_v_shift);
+static av_always_inline int check_4block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, int ref, int *best_rd){
+    const int b_stride= s->b_width << s->block_max_depth;
+    BlockNode *block= &s->block[mb_x + mb_y * b_stride];
+    BlockNode backup[4]= {block[0], block[1], block[b_stride], block[b_stride+1]};
+    int rd, index, value;
 
-    width= s->avctx->width;
-    height= s->avctx->height;
+    assert(mb_x>=0 && mb_y>=0);
+    assert(mb_x<b_stride);
+    assert(((mb_x|mb_y)&1) == 0);
 
-    s->spatial_idwt_buffer= av_mallocz(width*height*sizeof(IDWTELEM));
-    s->spatial_dwt_buffer= av_mallocz(width*height*sizeof(DWTELEM)); //FIXME this does not belong here
+    index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
+    value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
+    if(s->me_cache[index] == value)
+        return 0;
+    s->me_cache[index]= value;
 
-    for(i=0; i<MAX_REF_FRAMES; i++)
-        for(j=0; j<MAX_REF_FRAMES; j++)
-            scale_mv_ref[i][j] = 256*(i+1)/(j+1);
+    block->mx= p0;
+    block->my= p1;
+    block->ref= ref;
+    block->type &= ~BLOCK_INTRA;
+    block[1]= block[b_stride]= block[b_stride+1]= *block;
 
-    s->avctx->get_buffer(s->avctx, &s->mconly_picture);
-    s->scratchbuf = av_malloc(s->mconly_picture.linesize[0]*7*MB_SIZE);
+    rd= get_4block_rd(s, mb_x, mb_y, 0);
 
-    return 0;
+//FIXME chroma
+    if(rd < *best_rd){
+        *best_rd= rd;
+        return 1;
+    }else{
+        block[0]= backup[0];
+        block[1]= backup[1];
+        block[b_stride]= backup[2];
+        block[b_stride+1]= backup[3];
+        return 0;
+    }
 }
 
-static int common_init_after_header(AVCodecContext *avctx){
-    SnowContext *s = avctx->priv_data;
-    int plane_index, level, orientation;
-
-    for(plane_index=0; plane_index<3; plane_index++){
-        int w= s->avctx->width;
-        int h= s->avctx->height;
-
-        if(plane_index){
-            w>>= s->chroma_h_shift;
-            h>>= s->chroma_v_shift;
-        }
-        s->plane[plane_index].width = w;
-        s->plane[plane_index].height= h;
-
-        for(level=s->spatial_decomposition_count-1; level>=0; level--){
-            for(orientation=level ? 1 : 0; orientation<4; orientation++){
-                SubBand *b= &s->plane[plane_index].band[level][orientation];
+static void iterative_me(SnowContext *s){
+    int pass, mb_x, mb_y;
+    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;
+    int color[3];
 
-                b->buf= s->spatial_dwt_buffer;
-                b->level= level;
-                b->stride= s->plane[plane_index].width << (s->spatial_decomposition_count - level);
-                b->width = (w + !(orientation&1))>>1;
-                b->height= (h + !(orientation>1))>>1;
+    {
+        RangeCoder r = s->c;
+        uint8_t state[sizeof(s->block_state)];
+        memcpy(state, s->block_state, sizeof(s->block_state));
+        for(mb_y= 0; mb_y<s->b_height; mb_y++)
+            for(mb_x= 0; mb_x<s->b_width; mb_x++)
+                encode_q_branch(s, 0, mb_x, mb_y);
+        s->c = r;
+        memcpy(s->block_state, state, sizeof(s->block_state));
+    }
 
-                b->stride_line = 1 << (s->spatial_decomposition_count - level);
-                b->buf_x_offset = 0;
-                b->buf_y_offset = 0;
+    for(pass=0; pass<25; pass++){
+        int change= 0;
 
-                if(orientation&1){
-                    b->buf += (w+1)>>1;
-                    b->buf_x_offset = (w+1)>>1;
-                }
-                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);
+        for(mb_y= 0; mb_y<b_height; mb_y++){
+            for(mb_x= 0; mb_x<b_width; mb_x++){
+                int dia_change, i, j, ref;
+                int best_rd= INT_MAX, ref_rd;
+                BlockNode backup, ref_b;
+                const int index= mb_x + mb_y * b_stride;
+                BlockNode *block= &s->block[index];
+                BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
+                BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
+                BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
+                BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
+                BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
+                BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
+                BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
+                BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
+                const int b_w= (MB_SIZE >> s->block_max_depth);
+                uint8_t obmc_edged[b_w*2][b_w*2];
 
-                if(level)
-                    b->parent= &s->plane[plane_index].band[level-1][orientation];
-                //FIXME avoid this realloc
-                av_freep(&b->x_coeff);
-                b->x_coeff=av_mallocz(((b->width+1) * b->height+1)*sizeof(x_and_coeff));
-            }
-            w= (w+1)>>1;
-            h= (h+1)>>1;
-        }
-    }
+                if(pass && (block->type & BLOCK_OPT))
+                    continue;
+                block->type |= BLOCK_OPT;
 
-    return 0;
-}
+                backup= *block;
 
-static int qscale2qlog(int qscale){
-    return rint(QROOT*log(qscale / (float)FF_QP2LAMBDA)/log(2))
-           + 61*QROOT/8; //<64 >60
-}
+                if(!s->me_cache_generation)
+                    memset(s->me_cache, 0, sizeof(s->me_cache));
+                s->me_cache_generation += 1<<22;
 
-static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
-{
-    /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
-     * FIXME we know exact mv bits at this point,
-     * but ratecontrol isn't set up to include them. */
-    uint32_t coef_sum= 0;
-    int level, orientation, delta_qlog;
+                //FIXME precalculate
+                {
+                    int x, y;
+                    memcpy(obmc_edged, obmc_tab[s->block_max_depth], b_w*b_w*4);
+                    if(mb_x==0)
+                        for(y=0; y<b_w*2; y++)
+                            memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
+                    if(mb_x==b_stride-1)
+                        for(y=0; y<b_w*2; y++)
+                            memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
+                    if(mb_y==0){
+                        for(x=0; x<b_w*2; x++)
+                            obmc_edged[0][x] += obmc_edged[b_w-1][x];
+                        for(y=1; y<b_w; y++)
+                            memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
+                    }
+                    if(mb_y==b_height-1){
+                        for(x=0; x<b_w*2; x++)
+                            obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
+                        for(y=b_w; y<b_w*2-1; y++)
+                            memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
+                    }
+                }
 
-    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;
-        }
-    }
+                //skip stuff outside the picture
+                if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
+                    uint8_t *src= s->  input_picture.data[0];
+                    uint8_t *dst= s->current_picture.data[0];
+                    const int stride= s->current_picture.linesize[0];
+                    const int block_w= MB_SIZE >> s->block_max_depth;
+                    const int sx= block_w*mb_x - block_w/2;
+                    const int sy= block_w*mb_y - block_w/2;
+                    const int w= s->plane[0].width;
+                    const int h= s->plane[0].height;
+                    int y;
 
-    /* ugly, ratecontrol just takes a sqrt again */
-    coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
-    assert(coef_sum < INT_MAX);
+                    for(y=sy; y<0; y++)
+                        memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
+                    for(y=h; y<sy+block_w*2; y++)
+                        memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
+                    if(sx<0){
+                        for(y=sy; y<sy+block_w*2; y++)
+                            memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
+                    }
+                    if(sx+block_w*2 > w){
+                        for(y=sy; y<sy+block_w*2; y++)
+                            memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
+                    }
+                }
 
-    if(pict->pict_type == FF_I_TYPE){
-        s->m.current_picture.mb_var_sum= coef_sum;
-        s->m.current_picture.mc_mb_var_sum= 0;
-    }else{
-        s->m.current_picture.mc_mb_var_sum= coef_sum;
-        s->m.current_picture.mb_var_sum= 0;
-    }
+                // intra(black) = neighbors' contribution to the current block
+                for(i=0; i<3; i++)
+                    color[i]= get_dc(s, mb_x, mb_y, i);
 
-    pict->quality= ff_rate_estimate_qscale(&s->m, 1);
-    if (pict->quality < 0)
-        return INT_MIN;
-    s->lambda= pict->quality * 3/2;
-    delta_qlog= qscale2qlog(pict->quality) - s->qlog;
-    s->qlog+= delta_qlog;
-    return delta_qlog;
-}
+                // get previous score (cannot be cached due to OBMC)
+                if(pass > 0 && (block->type&BLOCK_INTRA)){
+                    int color0[3]= {block->color[0], block->color[1], block->color[2]};
+                    check_block(s, mb_x, mb_y, color0, 1, *obmc_edged, &best_rd);
+                }else
+                    check_block_inter(s, mb_x, mb_y, block->mx, block->my, *obmc_edged, &best_rd);
 
-static void calculate_visual_weight(SnowContext *s, Plane *p){
-    int width = p->width;
-    int height= p->height;
-    int level, orientation, x, y;
+                ref_b= *block;
+                ref_rd= best_rd;
+                for(ref=0; ref < s->ref_frames; ref++){
+                    int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
+                    if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
+                        continue;
+                    block->ref= ref;
+                    best_rd= INT_MAX;
 
-    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;
+                    check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], *obmc_edged, &best_rd);
+                    check_block_inter(s, mb_x, mb_y, 0, 0, *obmc_edged, &best_rd);
+                    if(tb)
+                        check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], *obmc_edged, &best_rd);
+                    if(lb)
+                        check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], *obmc_edged, &best_rd);
+                    if(rb)
+                        check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], *obmc_edged, &best_rd);
+                    if(bb)
+                        check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], *obmc_edged, &best_rd);
 
-            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;
+                    /* fullpel ME */
+                    //FIXME avoid subpel interpolation / round to nearest integer
+                    do{
+                        dia_change=0;
+                        for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
+                            for(j=0; j<i; j++){
+                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), *obmc_edged, &best_rd);
+                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), *obmc_edged, &best_rd);
+                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), *obmc_edged, &best_rd);
+                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), *obmc_edged, &best_rd);
+                            }
+                        }
+                    }while(dia_change);
+                    /* subpel ME */
+                    do{
+                        static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
+                        dia_change=0;
+                        for(i=0; i<8; i++)
+                            dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], *obmc_edged, &best_rd);
+                    }while(dia_change);
+                    //FIXME or try the standard 2 pass qpel or similar
+
+                    mvr[0][0]= block->mx;
+                    mvr[0][1]= block->my;
+                    if(ref_rd > best_rd){
+                        ref_rd= best_rd;
+                        ref_b= *block;
+                    }
+                }
+                best_rd= ref_rd;
+                *block= ref_b;
+                check_block(s, mb_x, mb_y, color, 1, *obmc_edged, &best_rd);
+                //FIXME RD style color selection
+                if(!same_block(block, &backup)){
+                    if(tb ) tb ->type &= ~BLOCK_OPT;
+                    if(lb ) lb ->type &= ~BLOCK_OPT;
+                    if(rb ) rb ->type &= ~BLOCK_OPT;
+                    if(bb ) bb ->type &= ~BLOCK_OPT;
+                    if(tlb) tlb->type &= ~BLOCK_OPT;
+                    if(trb) trb->type &= ~BLOCK_OPT;
+                    if(blb) blb->type &= ~BLOCK_OPT;
+                    if(brb) brb->type &= ~BLOCK_OPT;
+                    change ++;
                 }
             }
-
-            b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
         }
+        av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
+        if(!change)
+            break;
     }
-}
 
-#define QUANTIZE2 0
+    if(s->block_max_depth == 1){
+        int change= 0;
+        for(mb_y= 0; mb_y<b_height; mb_y+=2){
+            for(mb_x= 0; mb_x<b_width; mb_x+=2){
+                int i;
+                int best_rd, init_rd;
+                const int index= mb_x + mb_y * b_stride;
+                BlockNode *b[4];
+
+                b[0]= &s->block[index];
+                b[1]= b[0]+1;
+                b[2]= b[0]+b_stride;
+                b[3]= b[2]+1;
+                if(same_block(b[0], b[1]) &&
+                   same_block(b[0], b[2]) &&
+                   same_block(b[0], b[3]))
+                    continue;
+
+                if(!s->me_cache_generation)
+                    memset(s->me_cache, 0, sizeof(s->me_cache));
+                s->me_cache_generation += 1<<22;
 
-#if QUANTIZE2==1
-#define Q2_STEP 8
+                init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
 
-static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
-    SubBand *b= &p->band[level][orientation];
-    int x, y;
-    int xo=0;
-    int yo=0;
-    int step= 1 << (s->spatial_decomposition_count - level);
+                //FIXME more multiref search?
+                check_4block_inter(s, mb_x, mb_y,
+                                   (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
+                                   (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
 
-    if(orientation&1)
-        xo= step>>1;
-    if(orientation&2)
-        yo= step>>1;
+                for(i=0; i<4; i++)
+                    if(!(b[i]->type&BLOCK_INTRA))
+                        check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
 
-    //FIXME bias for nonzero ?
-    //FIXME optimize
-    memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
-    for(y=0; y<p->height; y++){
-        for(x=0; x<p->width; x++){
-            int sx= (x-xo + step/2) / step / Q2_STEP;
-            int sy= (y-yo + step/2) / step / Q2_STEP;
-            int v= r0[x + y*p->width] - r1[x + y*p->width];
-            assert(sx>=0 && sy>=0 && sx < score_stride);
-            v= ((v+8)>>4)<<4;
-            score[sx + sy*score_stride] += v*v;
-            assert(score[sx + sy*score_stride] >= 0);
+                if(init_rd != best_rd)
+                    change++;
+            }
         }
+        av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
     }
 }
 
-static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
-    int level, orientation;
+static void encode_blocks(SnowContext *s, int search){
+    int x, y;
+    int w= s->b_width;
+    int h= s->b_height;
 
-    for(level=0; level<s->spatial_decomposition_count; level++){
-        for(orientation=level ? 1 : 0; orientation<4; orientation++){
-            SubBand *b= &p->band[level][orientation];
-            IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
+    if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
+        iterative_me(s);
 
-            dequantize(s, b, dst, b->stride);
+    for(y=0; y<h; y++){
+        if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
+            av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
+            return;
+        }
+        for(x=0; x<w; x++){
+            if(s->avctx->me_method == ME_ITER || !search)
+                encode_q_branch2(s, 0, x, y);
+            else
+                encode_q_branch (s, 0, x, y);
         }
     }
 }
 
-static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
-    int level, orientation, ys, xs, x, y, pass;
-    IDWTELEM best_dequant[height * stride];
-    IDWTELEM idwt2_buffer[height * stride];
-    const int score_stride= (width + 10)/Q2_STEP;
-    int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
-    int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
-    int threshold= (s->m.lambda * s->m.lambda) >> 6;
+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;
 
-    //FIXME pass the copy cleanly ?
+    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;
+    }
 
-//    memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
-    ff_spatial_dwt(buffer, width, height, stride, type, s->spatial_decomposition_count);
+    bias= bias ? 0 : (3*qmul)>>3;
+    thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
+    thres2= 2*thres1;
 
-    for(level=0; level<s->spatial_decomposition_count; level++){
-        for(orientation=level ? 1 : 0; orientation<4; orientation++){
-            SubBand *b= &p->band[level][orientation];
-            IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
-             DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
-            assert(src == b->buf); // code does not depend on this but it is true currently
+    if(!bias){
+        for(y=0; y<h; y++){
+            for(x=0; x<w; x++){
+                int i= src[x + y*stride];
 
-            quantize(s, b, dst, src, b->stride, s->qbias);
+                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;
+            }
         }
-    }
-    for(pass=0; pass<1; pass++){
-        if(s->qbias == 0) //keyframe
-            continue;
-        for(level=0; level<s->spatial_decomposition_count; level++){
-            for(orientation=level ? 1 : 0; orientation<4; orientation++){
-                SubBand *b= &p->band[level][orientation];
-                IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
-                IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
+    }else{
+        for(y=0; y<h; y++){
+            for(x=0; x<w; x++){
+                int i= src[x + y*stride];
 
-                for(ys= 0; ys<Q2_STEP; ys++){
-                    for(xs= 0; xs<Q2_STEP; xs++){
-                        memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
-                        dequantize_all(s, p, idwt2_buffer, width, height);
-                        ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
-                        find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
-                        memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
-                        for(y=ys; y<b->height; y+= Q2_STEP){
-                            for(x=xs; x<b->width; x+= Q2_STEP){
-                                if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
-                                if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
-                                //FIXME try more than just --
-                            }
-                        }
-                        dequantize_all(s, p, idwt2_buffer, width, height);
-                        ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
-                        find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
-                        for(y=ys; y<b->height; y+= Q2_STEP){
-                            for(x=xs; x<b->width; x+= Q2_STEP){
-                                int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
-                                if(score[score_idx] <= best_score[score_idx] + threshold){
-                                    best_score[score_idx]= score[score_idx];
-                                    if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
-                                    if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
-                                    //FIXME copy instead
-                                }
-                            }
-                        }
+                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;
             }
         }
     }
-    memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
 }
 
-#endif /* QUANTIZE2==1 */
-
-static av_cold int encode_init(AVCodecContext *avctx)
-{
-    SnowContext *s = avctx->priv_data;
-    int plane_index;
-
-    if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
-        av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
-               "Use vstrict=-2 / -strict -2 to use it anyway.\n");
-        return -1;
-    }
-
-    if(avctx->prediction_method == DWT_97
-       && (avctx->flags & CODEC_FLAG_QSCALE)
-       && avctx->global_quality == 0){
-        av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
-        return -1;
-    }
-
-    s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
-
-    s->chroma_h_shift= 1; //FIXME XXX
-    s->chroma_v_shift= 1;
-
-    s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
-    s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
-
-    for(plane_index=0; plane_index<3; plane_index++){
-        s->plane[plane_index].diag_mc= 1;
-        s->plane[plane_index].htaps= 6;
-        s->plane[plane_index].hcoeff[0]=  40;
-        s->plane[plane_index].hcoeff[1]= -10;
-        s->plane[plane_index].hcoeff[2]=   2;
-        s->plane[plane_index].fast_mc= 1;
-    }
-
-    common_init(avctx);
-    alloc_blocks(s);
-
-    s->version=0;
-
-    s->m.avctx   = avctx;
-    s->m.flags   = avctx->flags;
-    s->m.bit_rate= avctx->bit_rate;
-
-    s->m.me.temp      =
-    s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
-    s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
-    s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
-    s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
-    h263_encode_init(&s->m); //mv_penalty
-
-    s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
-
-    if(avctx->flags&CODEC_FLAG_PASS1){
-        if(!avctx->stats_out)
-            avctx->stats_out = av_mallocz(256);
-    }
-    if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
-        if(ff_rate_control_init(&s->m) < 0)
-            return -1;
-    }
-    s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
-
-    avctx->coded_frame= &s->current_picture;
-    switch(avctx->pix_fmt){
-//    case PIX_FMT_YUV444P:
-//    case PIX_FMT_YUV422P:
-    case PIX_FMT_YUV420P:
-    case PIX_FMT_GRAY8:
-//    case PIX_FMT_YUV411P:
-//    case PIX_FMT_YUV410P:
-        s->colorspace_type= 0;
-        break;
-/*    case PIX_FMT_RGB32:
-        s->colorspace= 1;
-        break;*/
-    default:
-        av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
-        return -1;
-    }
-//    avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
-    s->chroma_h_shift= 1;
-    s->chroma_v_shift= 1;
-
-    ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
-    ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
+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;
 
-    s->avctx->get_buffer(s->avctx, &s->input_picture);
+    if(s->qlog == LOSSLESS_QLOG) return;
 
-    if(s->avctx->me_method == ME_ITER){
-        int i;
-        int size= s->b_width * s->b_height << 2*s->block_max_depth;
-        for(i=0; i<s->max_ref_frames; i++){
-            s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
-            s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
+    for(y=0; y<h; y++){
+        for(x=0; x<w; x++){
+            int i= src[x + y*stride];
+            if(i<0){
+                src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
+            }else if(i>0){
+                src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
+            }
         }
     }
-
-    return 0;
 }
 
-#define USE_HALFPEL_PLANE 0
-
-static void halfpel_interpol(SnowContext *s, uint8_t *halfpel[4][4], AVFrame *frame){
-    int p,x,y;
+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;
 
-    assert(!(s->avctx->flags & CODEC_FLAG_EMU_EDGE));
+    for(y=h-1; y>=0; y--){
+        for(x=w-1; x>=0; x--){
+            int i= x + y*stride;
 
-    for(p=0; p<3; p++){
-        int is_chroma= !!p;
-        int w= s->avctx->width  >>is_chroma;
-        int h= s->avctx->height >>is_chroma;
-        int ls= frame->linesize[p];
-        uint8_t *src= frame->data[p];
+            if(x){
+                if(use_median){
+                    if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
+                    else  src[i] -= src[i - 1];
+                }else{
+                    if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
+                    else  src[i] -= src[i - 1];
+                }
+            }else{
+                if(y) src[i] -= src[i - stride];
+            }
+        }
+    }
+}
 
-        halfpel[1][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls);
-        halfpel[2][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls);
-        halfpel[3][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls);
+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;
 
-        halfpel[0][p]= src;
-        for(y=0; y<h; y++){
-            for(x=0; x<w; x++){
-                int i= y*ls + x;
+    for(y=0; y<h; y++){
+        for(x=0; x<w; x++){
+            int i= x + y*stride;
 
-                halfpel[1][p][i]= (20*(src[i] + src[i+1]) - 5*(src[i-1] + src[i+2]) + (src[i-2] + src[i+3]) + 16 )>>5;
+            if(x){
+                if(use_median){
+                    if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
+                    else  src[i] += src[i - 1];
+                }else{
+                    if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
+                    else  src[i] += src[i - 1];
+                }
+            }else{
+                if(y) src[i] += src[i - stride];
             }
         }
-        for(y=0; y<h; y++){
-            for(x=0; x<w; x++){
-                int i= y*ls + x;
+    }
+}
 
-                halfpel[2][p][i]= (20*(src[i] + src[i+ls]) - 5*(src[i-ls] + src[i+2*ls]) + (src[i-2*ls] + src[i+3*ls]) + 16 )>>5;
+static void encode_qlogs(SnowContext *s){
+    int plane_index, level, orientation;
+
+    for(plane_index=0; plane_index<2; plane_index++){
+        for(level=0; level<s->spatial_decomposition_count; level++){
+            for(orientation=level ? 1:0; orientation<4; orientation++){
+                if(orientation==2) continue;
+                put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
             }
         }
-        src= halfpel[1][p];
-        for(y=0; y<h; y++){
-            for(x=0; x<w; x++){
-                int i= y*ls + x;
+    }
+}
 
-                halfpel[3][p][i]= (20*(src[i] + src[i+ls]) - 5*(src[i-ls] + src[i+2*ls]) + (src[i-2*ls] + src[i+3*ls]) + 16 )>>5;
+static void encode_header(SnowContext *s){
+    int plane_index, i;
+    uint8_t kstate[32];
+
+    memset(kstate, MID_STATE, sizeof(kstate));
+
+    put_rac(&s->c, kstate, s->keyframe);
+    if(s->keyframe || s->always_reset){
+        reset_contexts(s);
+        s->last_spatial_decomposition_type=
+        s->last_qlog=
+        s->last_qbias=
+        s->last_mv_scale=
+        s->last_block_max_depth= 0;
+        for(plane_index=0; plane_index<2; plane_index++){
+            Plane *p= &s->plane[plane_index];
+            p->last_htaps=0;
+            p->last_diag_mc=0;
+            memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
+        }
+    }
+    if(s->keyframe){
+        put_symbol(&s->c, s->header_state, s->version, 0);
+        put_rac(&s->c, s->header_state, s->always_reset);
+        put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
+        put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
+        put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
+        put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
+        put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
+        put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
+        put_rac(&s->c, s->header_state, s->spatial_scalability);
+//        put_rac(&s->c, s->header_state, s->rate_scalability);
+        put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
+
+        encode_qlogs(s);
+    }
+
+    if(!s->keyframe){
+        int update_mc=0;
+        for(plane_index=0; plane_index<2; plane_index++){
+            Plane *p= &s->plane[plane_index];
+            update_mc |= p->last_htaps   != p->htaps;
+            update_mc |= p->last_diag_mc != p->diag_mc;
+            update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
+        }
+        put_rac(&s->c, s->header_state, update_mc);
+        if(update_mc){
+            for(plane_index=0; plane_index<2; plane_index++){
+                Plane *p= &s->plane[plane_index];
+                put_rac(&s->c, s->header_state, p->diag_mc);
+                put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
+                for(i= p->htaps/2; i; i--)
+                    put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
             }
         }
+        if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
+            put_rac(&s->c, s->header_state, 1);
+            put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
+            encode_qlogs(s);
+        }else
+            put_rac(&s->c, s->header_state, 0);
+    }
 
-//FIXME border!
+    put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
+    put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
+    put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
+    put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
+    put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
+
+}
+
+static void update_last_header_values(SnowContext *s){
+    int plane_index;
+
+    if(!s->keyframe){
+        for(plane_index=0; plane_index<2; plane_index++){
+            Plane *p= &s->plane[plane_index];
+            p->last_diag_mc= p->diag_mc;
+            p->last_htaps  = p->htaps;
+            memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
+        }
     }
+
+    s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
+    s->last_qlog                        = s->qlog;
+    s->last_qbias                       = s->qbias;
+    s->last_mv_scale                    = s->mv_scale;
+    s->last_block_max_depth             = s->block_max_depth;
+    s->last_spatial_decomposition_count = s->spatial_decomposition_count;
 }
 
-static int frame_start(SnowContext *s){
-   AVFrame tmp;
-   int w= s->avctx->width; //FIXME round up to x16 ?
-   int h= s->avctx->height;
+static int qscale2qlog(int qscale){
+    return rint(QROOT*log(qscale / (float)FF_QP2LAMBDA)/log(2))
+           + 61*QROOT/8; //<64 >60
+}
 
-    if(s->current_picture.data[0]){
-        s->dsp.draw_edges(s->current_picture.data[0], s->current_picture.linesize[0], w   , h   , EDGE_WIDTH  );
-        s->dsp.draw_edges(s->current_picture.data[1], s->current_picture.linesize[1], w>>1, h>>1, EDGE_WIDTH/2);
-        s->dsp.draw_edges(s->current_picture.data[2], s->current_picture.linesize[2], w>>1, h>>1, EDGE_WIDTH/2);
+static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
+{
+    /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
+     * FIXME we know exact mv bits at this point,
+     * but ratecontrol isn't set up to include them. */
+    uint32_t coef_sum= 0;
+    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;
+        }
     }
 
-    tmp= s->last_picture[s->max_ref_frames-1];
-    memmove(s->last_picture+1, s->last_picture, (s->max_ref_frames-1)*sizeof(AVFrame));
-    memmove(s->halfpel_plane+1, s->halfpel_plane, (s->max_ref_frames-1)*sizeof(void*)*4*4);
-    if(USE_HALFPEL_PLANE && s->current_picture.data[0])
-        halfpel_interpol(s, s->halfpel_plane[0], &s->current_picture);
-    s->last_picture[0]= s->current_picture;
-    s->current_picture= tmp;
+    /* ugly, ratecontrol just takes a sqrt again */
+    coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
+    assert(coef_sum < INT_MAX);
 
-    if(s->keyframe){
-        s->ref_frames= 0;
+    if(pict->pict_type == AV_PICTURE_TYPE_I){
+        s->m.current_picture.mb_var_sum= coef_sum;
+        s->m.current_picture.mc_mb_var_sum= 0;
     }else{
-        int i;
-        for(i=0; i<s->max_ref_frames && s->last_picture[i].data[0]; i++)
-            if(i && s->last_picture[i-1].key_frame)
-                break;
-        s->ref_frames= i;
+        s->m.current_picture.mc_mb_var_sum= coef_sum;
+        s->m.current_picture.mb_var_sum= 0;
     }
 
-    s->current_picture.reference= 1;
-    if(s->avctx->get_buffer(s->avctx, &s->current_picture) < 0){
-        av_log(s->avctx, AV_LOG_ERROR, "get_buffer() failed\n");
-        return -1;
-    }
+    pict->quality= ff_rate_estimate_qscale(&s->m, 1);
+    if (pict->quality < 0)
+        return INT_MIN;
+    s->lambda= pict->quality * 3/2;
+    delta_qlog= qscale2qlog(pict->quality) - s->qlog;
+    s->qlog+= delta_qlog;
+    return delta_qlog;
+}
+
+static void calculate_visual_weight(SnowContext *s, Plane *p){
+    int width = p->width;
+    int height= p->height;
+    int level, orientation, x, y;
 
-    s->current_picture.key_frame= s->keyframe;
+    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;
 
-    return 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);
+        }
+    }
 }
 
 static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data){
@@ -4183,7 +3720,7 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size,
     if(avctx->flags&CODEC_FLAG_PASS2){
         s->m.pict_type =
         pict->pict_type= s->m.rc_context.entry[avctx->frame_number].new_pict_type;
-        s->keyframe= pict->pict_type==FF_I_TYPE;
+        s->keyframe= pict->pict_type==AV_PICTURE_TYPE_I;
         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
             pict->quality= ff_rate_estimate_qscale(&s->m, 0);
             if (pict->quality < 0)
@@ -4192,7 +3729,7 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size,
     }else{
         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
         s->m.pict_type=
-        pict->pict_type= s->keyframe ? FF_I_TYPE : FF_P_TYPE;
+        pict->pict_type= s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
     }
 
     if(s->pass1_rc && avctx->frame_number == 0)
@@ -4209,7 +3746,9 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size,
     frame_start(s);
 
     s->m.current_picture_ptr= &s->m.current_picture;
-    if(pict->pict_type == FF_P_TYPE){
+    s->m.last_picture.pts= s->m.current_picture.pts;
+    s->m.current_picture.pts= pict->pts;
+    if(pict->pict_type == AV_PICTURE_TYPE_P){
         int block_width = (width +15)>>4;
         int block_height= (height+15)>>4;
         int stride= s->current_picture.linesize[0];
@@ -4258,13 +3797,13 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size,
 
 redo_frame:
 
-    if(pict->pict_type == FF_I_TYPE)
+    if(pict->pict_type == AV_PICTURE_TYPE_I)
         s->spatial_decomposition_count= 5;
     else
         s->spatial_decomposition_count= 5;
 
     s->m.pict_type = pict->pict_type;
-    s->qbias= pict->pict_type == FF_P_TYPE ? 2 : 0;
+    s->qbias= pict->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
 
     common_init_after_header(avctx);
 
@@ -4297,12 +3836,12 @@ redo_frame:
             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
 
             if(   plane_index==0
-               && pict->pict_type == FF_P_TYPE
+               && pict->pict_type == AV_PICTURE_TYPE_P
                && !(avctx->flags&CODEC_FLAG_PASS2)
                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
                 ff_init_range_encoder(c, buf, buf_size);
                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
-                pict->pict_type= FF_I_TYPE;
+                pict->pict_type= AV_PICTURE_TYPE_I;
                 s->keyframe=1;
                 s->current_picture.key_frame=1;
                 goto redo_frame;
@@ -4348,7 +3887,7 @@ redo_frame:
                     if(!QUANTIZE2)
                         quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
                     if(orientation==0)
-                        decorrelate(s, b, b->ibuf, b->stride, pict->pict_type == FF_P_TYPE, 0);
+                        decorrelate(s, b, b->ibuf, b->stride, pict->pict_type == AV_PICTURE_TYPE_P, 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)
@@ -4375,7 +3914,7 @@ redo_frame:
             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
         }else{
             //ME/MC only
-            if(pict->pict_type == FF_I_TYPE){
+            if(pict->pict_type == AV_PICTURE_TYPE_I){
                 for(y=0; y<h; y++){
                     for(x=0; x<w; x++){
                         s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
@@ -4405,12 +3944,7 @@ redo_frame:
 
     update_last_header_values(s);
 
-    if(s->last_picture[s->max_ref_frames-1].data[0]){
-        avctx->release_buffer(avctx, &s->last_picture[s->max_ref_frames-1]);
-        for(i=0; i<9; i++)
-            if(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3])
-                av_free(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3] - EDGE_WIDTH*(1+s->current_picture.linesize[i%3]));
-    }
+    release_buffer(avctx);
 
     s->current_picture.coded_picture_number = avctx->frame_number;
     s->current_picture.pict_type = pict->pict_type;
@@ -4437,252 +3971,21 @@ redo_frame:
     return ff_rac_terminate(c);
 }
 
-static av_cold void common_end(SnowContext *s){
-    int plane_index, level, orientation, i;
-
-    av_freep(&s->spatial_dwt_buffer);
-    av_freep(&s->spatial_idwt_buffer);
-
-    s->m.me.temp= NULL;
-    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);
-
-    av_freep(&s->block);
-    av_freep(&s->scratchbuf);
-
-    for(i=0; i<MAX_REF_FRAMES; i++){
-        av_freep(&s->ref_mvs[i]);
-        av_freep(&s->ref_scores[i]);
-        if(s->last_picture[i].data[0])
-            s->avctx->release_buffer(s->avctx, &s->last_picture[i]);
-    }
-
-    for(plane_index=0; plane_index<3; plane_index++){
-        for(level=s->spatial_decomposition_count-1; level>=0; level--){
-            for(orientation=level ? 1 : 0; orientation<4; orientation++){
-                SubBand *b= &s->plane[plane_index].band[level][orientation];
-
-                av_freep(&b->x_coeff);
-            }
-        }
-    }
-}
-
 static av_cold int encode_end(AVCodecContext *avctx)
 {
     SnowContext *s = avctx->priv_data;
 
     common_end(s);
+    if (s->input_picture.data[0])
+        avctx->release_buffer(avctx, &s->input_picture);
     av_free(avctx->stats_out);
 
     return 0;
 }
 
-static av_cold int decode_init(AVCodecContext *avctx)
-{
-    avctx->pix_fmt= PIX_FMT_YUV420P;
-
-    common_init(avctx);
-
-    return 0;
-}
-
-static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, AVPacket *avpkt){
-    const uint8_t *buf = avpkt->data;
-    int buf_size = avpkt->size;
-    SnowContext *s = avctx->priv_data;
-    RangeCoder * const c= &s->c;
-    int bytes_read;
-    AVFrame *picture = data;
-    int level, orientation, plane_index, i;
-
-    ff_init_range_decoder(c, buf, buf_size);
-    ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
-
-    s->current_picture.pict_type= FF_I_TYPE; //FIXME I vs. P
-    if(decode_header(s)<0)
-        return -1;
-    common_init_after_header(avctx);
-
-    // realloc slice buffer for the case that spatial_decomposition_count changed
-    slice_buffer_destroy(&s->sb);
-    slice_buffer_init(&s->sb, s->plane[0].height, (MB_SIZE >> s->block_max_depth) + s->spatial_decomposition_count * 8 + 1, s->plane[0].width, s->spatial_idwt_buffer);
-
-    for(plane_index=0; plane_index<3; plane_index++){
-        Plane *p= &s->plane[plane_index];
-        p->fast_mc= p->diag_mc && p->htaps==6 && p->hcoeff[0]==40
-                                              && p->hcoeff[1]==-10
-                                              && p->hcoeff[2]==2;
-    }
-
-    if(!s->block) alloc_blocks(s);
-
-    frame_start(s);
-    //keyframe flag duplication mess FIXME
-    if(avctx->debug&FF_DEBUG_PICT_INFO)
-        av_log(avctx, AV_LOG_ERROR, "keyframe:%d qlog:%d\n", s->keyframe, s->qlog);
-
-    decode_blocks(s);
-
-    for(plane_index=0; plane_index<3; plane_index++){
-        Plane *p= &s->plane[plane_index];
-        int w= p->width;
-        int h= p->height;
-        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;
-                }
-            }
-        }
-
-        {
-        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;
-        DWTCompose cs[MAX_DECOMPOSITIONS];
-        int yd=0, yq=0;
-        int y;
-        int end_y;
-
-        ff_spatial_idwt_buffered_init(cs, &s->sb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count);
-        for(mb_y=0; mb_y<=mb_h; mb_y++){
-
-            int slice_starty = block_w*mb_y;
-            int slice_h = block_w*(mb_y+1);
-            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;
-                    int our_mb_start = mb_y;
-                    int our_mb_end = (mb_y + 1);
-                    const int extra= 3;
-                    start_y = (mb_y ? ((block_w * our_mb_start) >> (s->spatial_decomposition_count - level)) + s->spatial_decomposition_count - level + extra: 0);
-                    end_y = (((block_w * our_mb_end) >> (s->spatial_decomposition_count - level)) + s->spatial_decomposition_count - level + extra);
-                    if (!(s->keyframe || s->avctx->debug&512)){
-                        start_y = FFMAX(0, start_y - (block_w >> (1+s->spatial_decomposition_count - level)));
-                        end_y = FFMAX(0, end_y - (block_w >> (1+s->spatial_decomposition_count - level)));
-                    }
-                    start_y = FFMIN(b->height, start_y);
-                    end_y = FFMIN(b->height, end_y);
-
-                    if (start_y != end_y){
-                        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]);
-                    }
-                }
-            }
-
-            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++){
-                        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++);
-        }
-
-        slice_buffer_flush(&s->sb);
-        }
-
-    }
-
-    emms_c();
-
-    if(s->last_picture[s->max_ref_frames-1].data[0]){
-        avctx->release_buffer(avctx, &s->last_picture[s->max_ref_frames-1]);
-        for(i=0; i<9; i++)
-            if(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3])
-                av_free(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3] - EDGE_WIDTH*(1+s->current_picture.linesize[i%3]));
-    }
-
-    if(!(s->avctx->debug&2048))
-        *picture= s->current_picture;
-    else
-        *picture= s->mconly_picture;
-
-    *data_size = sizeof(AVFrame);
-
-    bytes_read= c->bytestream - c->bytestream_start;
-    if(bytes_read ==0) av_log(s->avctx, AV_LOG_ERROR, "error at end of frame\n"); //FIXME
-
-    return bytes_read;
-}
-
-static av_cold int decode_end(AVCodecContext *avctx)
-{
-    SnowContext *s = avctx->priv_data;
-
-    slice_buffer_destroy(&s->sb);
-
-    common_end(s);
-
-    return 0;
-}
-
-AVCodec snow_decoder = {
-    "snow",
-    CODEC_TYPE_VIDEO,
-    CODEC_ID_SNOW,
-    sizeof(SnowContext),
-    decode_init,
-    NULL,
-    decode_end,
-    decode_frame,
-    0 /*CODEC_CAP_DR1*/ /*| CODEC_CAP_DRAW_HORIZ_BAND*/,
-    NULL,
-    .long_name = NULL_IF_CONFIG_SMALL("Snow"),
-};
-
-#if CONFIG_SNOW_ENCODER
-AVCodec snow_encoder = {
+AVCodec ff_snow_encoder = {
     "snow",
-    CODEC_TYPE_VIDEO,
+    AVMEDIA_TYPE_VIDEO,
     CODEC_ID_SNOW,
     sizeof(SnowContext),
     encode_init,
@@ -4706,15 +4009,15 @@ int main(void){
     int buffer[2][width*height];
     SnowContext s;
     int i;
-    AVLFG prn;
+    AVLFG prng;
     s.spatial_decomposition_count=6;
     s.spatial_decomposition_type=1;
 
-    av_lfg_init(&prn, 1);
+    av_lfg_init(&prng, 1);
 
     printf("testing 5/3 DWT\n");
     for(i=0; i<width*height; i++)
-        buffer[0][i] = buffer[1][i] = av_lfg_get(&prn) % 54321 - 12345;
+        buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
 
     ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
     ff_spatial_idwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
@@ -4725,7 +4028,7 @@ int main(void){
     printf("testing 9/7 DWT\n");
     s.spatial_decomposition_type=0;
     for(i=0; i<width*height; i++)
-        buffer[0][i] = buffer[1][i] = av_lfg_get(&prn) % 54321 - 12345;
+        buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
 
     ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
     ff_spatial_idwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);