yading@10: /*
yading@10:  * Motion estimation
yading@10:  * Copyright (c) 2002-2004 Michael Niedermayer
yading@10:  *
yading@10:  * This file is part of FFmpeg.
yading@10:  *
yading@10:  * FFmpeg is free software; you can redistribute it and/or
yading@10:  * modify it under the terms of the GNU Lesser General Public
yading@10:  * License as published by the Free Software Foundation; either
yading@10:  * version 2.1 of the License, or (at your option) any later version.
yading@10:  *
yading@10:  * FFmpeg is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@10:  * Lesser General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU Lesser General Public
yading@10:  * License along with FFmpeg; if not, write to the Free Software
yading@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10:  */
yading@10: 
yading@10: /**
yading@10:  * @file
yading@10:  * Motion estimation template.
yading@10:  */
yading@10: 
yading@10: //Let us hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)
yading@10: #define LOAD_COMMON\
yading@10:     uint32_t av_unused * const score_map= c->score_map;\
yading@10:     const int av_unused xmin= c->xmin;\
yading@10:     const int av_unused ymin= c->ymin;\
yading@10:     const int av_unused xmax= c->xmax;\
yading@10:     const int av_unused ymax= c->ymax;\
yading@10:     uint8_t *mv_penalty= c->current_mv_penalty;\
yading@10:     const int pred_x= c->pred_x;\
yading@10:     const int pred_y= c->pred_y;\
yading@10: 
yading@10: #define CHECK_HALF_MV(dx, dy, x, y)\
yading@10: {\
yading@10:     const int hx= 2*(x)+(dx);\
yading@10:     const int hy= 2*(y)+(dy);\
yading@10:     d= cmp_hpel(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\
yading@10:     d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
yading@10:     COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
yading@10: }
yading@10: 
yading@10: static int hpel_motion_search(MpegEncContext * s,
yading@10:                                   int *mx_ptr, int *my_ptr, int dmin,
yading@10:                                   int src_index, int ref_index,
yading@10:                                   int size, int h)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     const int mx = *mx_ptr;
yading@10:     const int my = *my_ptr;
yading@10:     const int penalty_factor= c->sub_penalty_factor;
yading@10:     me_cmp_func cmp_sub, chroma_cmp_sub;
yading@10:     int bx=2*mx, by=2*my;
yading@10: 
yading@10:     LOAD_COMMON
yading@10:     int flags= c->sub_flags;
yading@10: 
yading@10:  //FIXME factorize
yading@10: 
yading@10:     cmp_sub= s->dsp.me_sub_cmp[size];
yading@10:     chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
yading@10: 
yading@10:     if(c->skip){ //FIXME move out of hpel?
yading@10:         *mx_ptr = 0;
yading@10:         *my_ptr = 0;
yading@10:         return dmin;
yading@10:     }
yading@10: 
yading@10:     if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
yading@10:         dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
yading@10:         if(mx || my || size>0)
yading@10:             dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
yading@10:     }
yading@10: 
yading@10:     if (mx > xmin && mx < xmax &&
yading@10:         my > ymin && my < ymax) {
yading@10:         int d= dmin;
yading@10:         const int index= (my<<ME_MAP_SHIFT) + mx;
yading@10:         const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
yading@10:                      + (mv_penalty[bx   - pred_x] + mv_penalty[by-2 - pred_y])*c->penalty_factor;
yading@10:         const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)]
yading@10:                      + (mv_penalty[bx-2 - pred_x] + mv_penalty[by   - pred_y])*c->penalty_factor;
yading@10:         const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)]
yading@10:                      + (mv_penalty[bx+2 - pred_x] + mv_penalty[by   - pred_y])*c->penalty_factor;
yading@10:         const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
yading@10:                      + (mv_penalty[bx   - pred_x] + mv_penalty[by+2 - pred_y])*c->penalty_factor;
yading@10: 
yading@10: #if defined(ASSERT_LEVEL) && ASSERT_LEVEL > 1
yading@10:         unsigned key;
yading@10:         unsigned map_generation= c->map_generation;
yading@10:         key= ((my-1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
yading@10:         av_assert2(c->map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
yading@10:         key= ((my+1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
yading@10:         av_assert2(c->map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
yading@10:         key= ((my)<<ME_MAP_MV_BITS) + (mx+1) + map_generation;
yading@10:         av_assert2(c->map[(index+1)&(ME_MAP_SIZE-1)] == key);
yading@10:         key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation;
yading@10:         av_assert2(c->map[(index-1)&(ME_MAP_SIZE-1)] == key);
yading@10: #endif
yading@10:         if(t<=b){
yading@10:             CHECK_HALF_MV(0, 1, mx  ,my-1)
yading@10:             if(l<=r){
yading@10:                 CHECK_HALF_MV(1, 1, mx-1, my-1)
yading@10:                 if(t+r<=b+l){
yading@10:                     CHECK_HALF_MV(1, 1, mx  , my-1)
yading@10:                 }else{
yading@10:                     CHECK_HALF_MV(1, 1, mx-1, my  )
yading@10:                 }
yading@10:                 CHECK_HALF_MV(1, 0, mx-1, my  )
yading@10:             }else{
yading@10:                 CHECK_HALF_MV(1, 1, mx  , my-1)
yading@10:                 if(t+l<=b+r){
yading@10:                     CHECK_HALF_MV(1, 1, mx-1, my-1)
yading@10:                 }else{
yading@10:                     CHECK_HALF_MV(1, 1, mx  , my  )
yading@10:                 }
yading@10:                 CHECK_HALF_MV(1, 0, mx  , my  )
yading@10:             }
yading@10:         }else{
yading@10:             if(l<=r){
yading@10:                 if(t+l<=b+r){
yading@10:                     CHECK_HALF_MV(1, 1, mx-1, my-1)
yading@10:                 }else{
yading@10:                     CHECK_HALF_MV(1, 1, mx  , my  )
yading@10:                 }
yading@10:                 CHECK_HALF_MV(1, 0, mx-1, my)
yading@10:                 CHECK_HALF_MV(1, 1, mx-1, my)
yading@10:             }else{
yading@10:                 if(t+r<=b+l){
yading@10:                     CHECK_HALF_MV(1, 1, mx  , my-1)
yading@10:                 }else{
yading@10:                     CHECK_HALF_MV(1, 1, mx-1, my)
yading@10:                 }
yading@10:                 CHECK_HALF_MV(1, 0, mx  , my)
yading@10:                 CHECK_HALF_MV(1, 1, mx  , my)
yading@10:             }
yading@10:             CHECK_HALF_MV(0, 1, mx  , my)
yading@10:         }
yading@10:         av_assert2(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2);
yading@10:     }
yading@10: 
yading@10:     *mx_ptr = bx;
yading@10:     *my_ptr = by;
yading@10: 
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: static int no_sub_motion_search(MpegEncContext * s,
yading@10:           int *mx_ptr, int *my_ptr, int dmin,
yading@10:                                   int src_index, int ref_index,
yading@10:                                   int size, int h)
yading@10: {
yading@10:     (*mx_ptr)<<=1;
yading@10:     (*my_ptr)<<=1;
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: static inline int get_mb_score(MpegEncContext *s, int mx, int my,
yading@10:                                int src_index, int ref_index, int size,
yading@10:                                int h, int add_rate)
yading@10: {
yading@10: //    const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp;
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     const int penalty_factor= c->mb_penalty_factor;
yading@10:     const int flags= c->mb_flags;
yading@10:     const int qpel= flags & FLAG_QPEL;
yading@10:     const int mask= 1+2*qpel;
yading@10:     me_cmp_func cmp_sub, chroma_cmp_sub;
yading@10:     int d;
yading@10: 
yading@10:     LOAD_COMMON
yading@10: 
yading@10:  //FIXME factorize
yading@10: 
yading@10:     cmp_sub= s->dsp.mb_cmp[size];
yading@10:     chroma_cmp_sub= s->dsp.mb_cmp[size+1];
yading@10: 
yading@10:     d= cmp(s, mx>>(qpel+1), my>>(qpel+1), mx&mask, my&mask, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
yading@10:     //FIXME check cbp before adding penalty for (0,0) vector
yading@10:     if(add_rate && (mx || my || size>0))
yading@10:         d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;
yading@10: 
yading@10:     return d;
yading@10: }
yading@10: 
yading@10: int ff_get_mb_score(MpegEncContext *s, int mx, int my, int src_index,
yading@10:                     int ref_index, int size, int h, int add_rate)
yading@10: {
yading@10:     return get_mb_score(s, mx, my, src_index, ref_index, size, h, add_rate);
yading@10: }
yading@10: 
yading@10: #define CHECK_QUARTER_MV(dx, dy, x, y)\
yading@10: {\
yading@10:     const int hx= 4*(x)+(dx);\
yading@10:     const int hy= 4*(y)+(dy);\
yading@10:     d= cmp_qpel(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
yading@10:     d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
yading@10:     COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
yading@10: }
yading@10: 
yading@10: static int qpel_motion_search(MpegEncContext * s,
yading@10:                                   int *mx_ptr, int *my_ptr, int dmin,
yading@10:                                   int src_index, int ref_index,
yading@10:                                   int size, int h)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     const int mx = *mx_ptr;
yading@10:     const int my = *my_ptr;
yading@10:     const int penalty_factor= c->sub_penalty_factor;
yading@10:     const unsigned map_generation = c->map_generation;
yading@10:     const int subpel_quality= c->avctx->me_subpel_quality;
yading@10:     uint32_t *map= c->map;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     me_cmp_func cmp_sub, chroma_cmp_sub;
yading@10: 
yading@10:     LOAD_COMMON
yading@10:     int flags= c->sub_flags;
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1]; //factorize FIXME
yading@10:  //FIXME factorize
yading@10: 
yading@10:     cmp_sub= s->dsp.me_sub_cmp[size];
yading@10:     chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
yading@10: 
yading@10:     if(c->skip){ //FIXME somehow move up (benchmark)
yading@10:         *mx_ptr = 0;
yading@10:         *my_ptr = 0;
yading@10:         return dmin;
yading@10:     }
yading@10: 
yading@10:     if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
yading@10:         dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
yading@10:         if(mx || my || size>0)
yading@10:             dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor;
yading@10:     }
yading@10: 
yading@10:     if (mx > xmin && mx < xmax &&
yading@10:         my > ymin && my < ymax) {
yading@10:         int bx=4*mx, by=4*my;
yading@10:         int d= dmin;
yading@10:         int i, nx, ny;
yading@10:         const int index= (my<<ME_MAP_SHIFT) + mx;
yading@10:         const int t= score_map[(index-(1<<ME_MAP_SHIFT)  )&(ME_MAP_SIZE-1)];
yading@10:         const int l= score_map[(index- 1                 )&(ME_MAP_SIZE-1)];
yading@10:         const int r= score_map[(index+ 1                 )&(ME_MAP_SIZE-1)];
yading@10:         const int b= score_map[(index+(1<<ME_MAP_SHIFT)  )&(ME_MAP_SIZE-1)];
yading@10:         const int c= score_map[(index                    )&(ME_MAP_SIZE-1)];
yading@10:         int best[8];
yading@10:         int best_pos[8][2];
yading@10: 
yading@10:         memset(best, 64, sizeof(int)*8);
yading@10:         if(s->me.dia_size>=2){
yading@10:             const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
yading@10:             const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
yading@10:             const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
yading@10:             const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
yading@10: 
yading@10:             for(ny= -3; ny <= 3; ny++){
yading@10:                 for(nx= -3; nx <= 3; nx++){
yading@10:                     //FIXME this could overflow (unlikely though)
yading@10:                     const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;
yading@10:                     const int64_t c2= nx*nx*( r +  l - 2*c) + 4*nx*( r- l) + 32*c;
yading@10:                     const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b;
yading@10:                     int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10;
yading@10:                     int i;
yading@10: 
yading@10:                     if((nx&3)==0 && (ny&3)==0) continue;
yading@10: 
yading@10:                     score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
yading@10: 
yading@10: //                    if(nx&1) score-=1024*c->penalty_factor;
yading@10: //                    if(ny&1) score-=1024*c->penalty_factor;
yading@10: 
yading@10:                     for(i=0; i<8; i++){
yading@10:                         if(score < best[i]){
yading@10:                             memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
yading@10:                             memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
yading@10:                             best[i]= score;
yading@10:                             best_pos[i][0]= nx + 4*mx;
yading@10:                             best_pos[i][1]= ny + 4*my;
yading@10:                             break;
yading@10:                         }
yading@10:                     }
yading@10:                 }
yading@10:             }
yading@10:         }else{
yading@10:             int tl;
yading@10:             //FIXME this could overflow (unlikely though)
yading@10:             const int cx = 4*(r - l);
yading@10:             const int cx2= r + l - 2*c;
yading@10:             const int cy = 4*(b - t);
yading@10:             const int cy2= b + t - 2*c;
yading@10:             int cxy;
yading@10: 
yading@10:             if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME
yading@10:                 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
yading@10:             }else{
yading@10:                 tl= cmp(s, mx-1, my-1, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);//FIXME wrong if chroma me is different
yading@10:             }
yading@10: 
yading@10:             cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;
yading@10: 
yading@10:             av_assert2(16*cx2 + 4*cx + 32*c == 32*r);
yading@10:             av_assert2(16*cx2 - 4*cx + 32*c == 32*l);
yading@10:             av_assert2(16*cy2 + 4*cy + 32*c == 32*b);
yading@10:             av_assert2(16*cy2 - 4*cy + 32*c == 32*t);
yading@10:             av_assert2(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);
yading@10: 
yading@10:             for(ny= -3; ny <= 3; ny++){
yading@10:                 for(nx= -3; nx <= 3; nx++){
yading@10:                     //FIXME this could overflow (unlikely though)
yading@10:                     int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor
yading@10:                     int i;
yading@10: 
yading@10:                     if((nx&3)==0 && (ny&3)==0) continue;
yading@10: 
yading@10:                     score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
yading@10: //                    if(nx&1) score-=32*c->penalty_factor;
yading@10:   //                  if(ny&1) score-=32*c->penalty_factor;
yading@10: 
yading@10:                     for(i=0; i<8; i++){
yading@10:                         if(score < best[i]){
yading@10:                             memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
yading@10:                             memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
yading@10:                             best[i]= score;
yading@10:                             best_pos[i][0]= nx + 4*mx;
yading@10:                             best_pos[i][1]= ny + 4*my;
yading@10:                             break;
yading@10:                         }
yading@10:                     }
yading@10:                 }
yading@10:             }
yading@10:         }
yading@10:         for(i=0; i<subpel_quality; i++){
yading@10:             nx= best_pos[i][0];
yading@10:             ny= best_pos[i][1];
yading@10:             CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
yading@10:         }
yading@10: 
yading@10:         av_assert2(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);
yading@10: 
yading@10:         *mx_ptr = bx;
yading@10:         *my_ptr = by;
yading@10:     }else{
yading@10:         *mx_ptr =4*mx;
yading@10:         *my_ptr =4*my;
yading@10:     }
yading@10: 
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: 
yading@10: #define CHECK_MV(x,y)\
yading@10: {\
yading@10:     const unsigned key = ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
yading@10:     const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
yading@10:     av_assert2((x) >= xmin);\
yading@10:     av_assert2((x) <= xmax);\
yading@10:     av_assert2((y) >= ymin);\
yading@10:     av_assert2((y) <= ymax);\
yading@10:     if(map[index]!=key){\
yading@10:         d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
yading@10:         map[index]= key;\
yading@10:         score_map[index]= d;\
yading@10:         d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
yading@10:         COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
yading@10:     }\
yading@10: }
yading@10: 
yading@10: #define CHECK_CLIPPED_MV(ax,ay)\
yading@10: {\
yading@10:     const int Lx= ax;\
yading@10:     const int Ly= ay;\
yading@10:     const int Lx2= FFMAX(xmin, FFMIN(Lx, xmax));\
yading@10:     const int Ly2= FFMAX(ymin, FFMIN(Ly, ymax));\
yading@10:     CHECK_MV(Lx2, Ly2)\
yading@10: }
yading@10: 
yading@10: #define CHECK_MV_DIR(x,y,new_dir)\
yading@10: {\
yading@10:     const unsigned key = ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
yading@10:     const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
yading@10:     if(map[index]!=key){\
yading@10:         d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
yading@10:         map[index]= key;\
yading@10:         score_map[index]= d;\
yading@10:         d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
yading@10:         if(d<dmin){\
yading@10:             best[0]=x;\
yading@10:             best[1]=y;\
yading@10:             dmin=d;\
yading@10:             next_dir= new_dir;\
yading@10:         }\
yading@10:     }\
yading@10: }
yading@10: 
yading@10: #define check(x,y,S,v)\
yading@10: if( (x)<(xmin<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\
yading@10: if( (x)>(xmax<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\
yading@10: if( (y)<(ymin<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\
yading@10: if( (y)>(ymax<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\
yading@10: 
yading@10: #define LOAD_COMMON2\
yading@10:     uint32_t *map= c->map;\
yading@10:     const int qpel= flags&FLAG_QPEL;\
yading@10:     const int shift= 1+qpel;\
yading@10: 
yading@10: static av_always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
yading@10:                                        int src_index, int ref_index, int const penalty_factor,
yading@10:                                        int size, int h, int flags)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     int next_dir=-1;
yading@10:     LOAD_COMMON
yading@10:     LOAD_COMMON2
yading@10:     unsigned map_generation = c->map_generation;
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     { /* ensure that the best point is in the MAP as h/qpel refinement needs it */
yading@10:         const unsigned key = (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;
yading@10:         const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);
yading@10:         if(map[index]!=key){ //this will be executed only very rarey
yading@10:             score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
yading@10:             map[index]= key;
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     for(;;){
yading@10:         int d;
yading@10:         const int dir= next_dir;
yading@10:         const int x= best[0];
yading@10:         const int y= best[1];
yading@10:         next_dir=-1;
yading@10: 
yading@10:         if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y  , 0)
yading@10:         if(dir!=3 && y>ymin) CHECK_MV_DIR(x  , y-1, 1)
yading@10:         if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y  , 2)
yading@10:         if(dir!=1 && y<ymax) CHECK_MV_DIR(x  , y+1, 3)
yading@10: 
yading@10:         if(next_dir==-1){
yading@10:             return dmin;
yading@10:         }
yading@10:     }
yading@10: }
yading@10: 
yading@10: static int funny_diamond_search(MpegEncContext * s, int *best, int dmin,
yading@10:                                        int src_index, int ref_index, int const penalty_factor,
yading@10:                                        int size, int h, int flags)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     int dia_size;
yading@10:     LOAD_COMMON
yading@10:     LOAD_COMMON2
yading@10:     unsigned map_generation = c->map_generation;
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     for(dia_size=1; dia_size<=4; dia_size++){
yading@10:         int dir;
yading@10:         const int x= best[0];
yading@10:         const int y= best[1];
yading@10: 
yading@10:         if(dia_size&(dia_size-1)) continue;
yading@10: 
yading@10:         if(   x + dia_size > xmax
yading@10:            || x - dia_size < xmin
yading@10:            || y + dia_size > ymax
yading@10:            || y - dia_size < ymin)
yading@10:            continue;
yading@10: 
yading@10:         for(dir= 0; dir<dia_size; dir+=2){
yading@10:             int d;
yading@10: 
yading@10:             CHECK_MV(x + dir           , y + dia_size - dir);
yading@10:             CHECK_MV(x + dia_size - dir, y - dir           );
yading@10:             CHECK_MV(x - dir           , y - dia_size + dir);
yading@10:             CHECK_MV(x - dia_size + dir, y + dir           );
yading@10:         }
yading@10: 
yading@10:         if(x!=best[0] || y!=best[1])
yading@10:             dia_size=0;
yading@10:     }
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: static int hex_search(MpegEncContext * s, int *best, int dmin,
yading@10:                                        int src_index, int ref_index, int const penalty_factor,
yading@10:                                        int size, int h, int flags, int dia_size)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     LOAD_COMMON
yading@10:     LOAD_COMMON2
yading@10:     unsigned map_generation = c->map_generation;
yading@10:     int x,y,d;
yading@10:     const int dec= dia_size & (dia_size-1);
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     for(;dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){
yading@10:         do{
yading@10:             x= best[0];
yading@10:             y= best[1];
yading@10: 
yading@10:             CHECK_CLIPPED_MV(x  -dia_size    , y);
yading@10:             CHECK_CLIPPED_MV(x+  dia_size    , y);
yading@10:             CHECK_CLIPPED_MV(x+( dia_size>>1), y+dia_size);
yading@10:             CHECK_CLIPPED_MV(x+( dia_size>>1), y-dia_size);
yading@10:             if(dia_size>1){
yading@10:                 CHECK_CLIPPED_MV(x+(-dia_size>>1), y+dia_size);
yading@10:                 CHECK_CLIPPED_MV(x+(-dia_size>>1), y-dia_size);
yading@10:             }
yading@10:         }while(best[0] != x || best[1] != y);
yading@10:     }
yading@10: 
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: static int l2s_dia_search(MpegEncContext * s, int *best, int dmin,
yading@10:                                        int src_index, int ref_index, int const penalty_factor,
yading@10:                                        int size, int h, int flags)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     LOAD_COMMON
yading@10:     LOAD_COMMON2
yading@10:     unsigned map_generation = c->map_generation;
yading@10:     int x,y,i,d;
yading@10:     int dia_size= c->dia_size&0xFF;
yading@10:     const int dec= dia_size & (dia_size-1);
yading@10:     static const int hex[8][2]={{-2, 0}, {-1,-1}, { 0,-2}, { 1,-1},
yading@10:                                 { 2, 0}, { 1, 1}, { 0, 2}, {-1, 1}};
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     for(; dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){
yading@10:         do{
yading@10:             x= best[0];
yading@10:             y= best[1];
yading@10:             for(i=0; i<8; i++){
yading@10:                 CHECK_CLIPPED_MV(x+hex[i][0]*dia_size, y+hex[i][1]*dia_size);
yading@10:             }
yading@10:         }while(best[0] != x || best[1] != y);
yading@10:     }
yading@10: 
yading@10:     x= best[0];
yading@10:     y= best[1];
yading@10:     CHECK_CLIPPED_MV(x+1, y);
yading@10:     CHECK_CLIPPED_MV(x, y+1);
yading@10:     CHECK_CLIPPED_MV(x-1, y);
yading@10:     CHECK_CLIPPED_MV(x, y-1);
yading@10: 
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: static int umh_search(MpegEncContext * s, int *best, int dmin,
yading@10:                                        int src_index, int ref_index, int const penalty_factor,
yading@10:                                        int size, int h, int flags)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     LOAD_COMMON
yading@10:     LOAD_COMMON2
yading@10:     unsigned map_generation = c->map_generation;
yading@10:     int x,y,x2,y2, i, j, d;
yading@10:     const int dia_size= c->dia_size&0xFE;
yading@10:     static const int hex[16][2]={{-4,-2}, {-4,-1}, {-4, 0}, {-4, 1}, {-4, 2},
yading@10:                                  { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2},
yading@10:                                  {-2, 3}, { 0, 4}, { 2, 3},
yading@10:                                  {-2,-3}, { 0,-4}, { 2,-3},};
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     x= best[0];
yading@10:     y= best[1];
yading@10:     for(x2=FFMAX(x-dia_size+1, xmin); x2<=FFMIN(x+dia_size-1,xmax); x2+=2){
yading@10:         CHECK_MV(x2, y);
yading@10:     }
yading@10:     for(y2=FFMAX(y-dia_size/2+1, ymin); y2<=FFMIN(y+dia_size/2-1,ymax); y2+=2){
yading@10:         CHECK_MV(x, y2);
yading@10:     }
yading@10: 
yading@10:     x= best[0];
yading@10:     y= best[1];
yading@10:     for(y2=FFMAX(y-2, ymin); y2<=FFMIN(y+2,ymax); y2++){
yading@10:         for(x2=FFMAX(x-2, xmin); x2<=FFMIN(x+2,xmax); x2++){
yading@10:             CHECK_MV(x2, y2);
yading@10:         }
yading@10:     }
yading@10: 
yading@10: //FIXME prevent the CLIP stuff
yading@10: 
yading@10:     for(j=1; j<=dia_size/4; j++){
yading@10:         for(i=0; i<16; i++){
yading@10:             CHECK_CLIPPED_MV(x+hex[i][0]*j, y+hex[i][1]*j);
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, 2);
yading@10: }
yading@10: 
yading@10: static int full_search(MpegEncContext * s, int *best, int dmin,
yading@10:                                        int src_index, int ref_index, int const penalty_factor,
yading@10:                                        int size, int h, int flags)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     LOAD_COMMON
yading@10:     LOAD_COMMON2
yading@10:     unsigned map_generation = c->map_generation;
yading@10:     int x,y, d;
yading@10:     const int dia_size= c->dia_size&0xFF;
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     for(y=FFMAX(-dia_size, ymin); y<=FFMIN(dia_size,ymax); y++){
yading@10:         for(x=FFMAX(-dia_size, xmin); x<=FFMIN(dia_size,xmax); x++){
yading@10:             CHECK_MV(x, y);
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     x= best[0];
yading@10:     y= best[1];
yading@10:     d= dmin;
yading@10:     CHECK_CLIPPED_MV(x  , y);
yading@10:     CHECK_CLIPPED_MV(x+1, y);
yading@10:     CHECK_CLIPPED_MV(x, y+1);
yading@10:     CHECK_CLIPPED_MV(x-1, y);
yading@10:     CHECK_CLIPPED_MV(x, y-1);
yading@10:     best[0]= x;
yading@10:     best[1]= y;
yading@10: 
yading@10:     return d;
yading@10: }
yading@10: 
yading@10: #define SAB_CHECK_MV(ax,ay)\
yading@10: {\
yading@10:     const unsigned key = ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\
yading@10:     const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\
yading@10:     if(map[index]!=key){\
yading@10:         d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
yading@10:         map[index]= key;\
yading@10:         score_map[index]= d;\
yading@10:         d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\
yading@10:         if(d < minima[minima_count-1].height){\
yading@10:             int j=0;\
yading@10:             \
yading@10:             while(d >= minima[j].height) j++;\
yading@10: \
yading@10:             memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\
yading@10: \
yading@10:             minima[j].checked= 0;\
yading@10:             minima[j].height= d;\
yading@10:             minima[j].x= ax;\
yading@10:             minima[j].y= ay;\
yading@10:             \
yading@10:             i=-1;\
yading@10:             continue;\
yading@10:         }\
yading@10:     }\
yading@10: }
yading@10: 
yading@10: #define MAX_SAB_SIZE ME_MAP_SIZE
yading@10: static int sab_diamond_search(MpegEncContext * s, int *best, int dmin,
yading@10:                                        int src_index, int ref_index, int const penalty_factor,
yading@10:                                        int size, int h, int flags)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     Minima minima[MAX_SAB_SIZE];
yading@10:     const int minima_count= FFABS(c->dia_size);
yading@10:     int i, j;
yading@10:     LOAD_COMMON
yading@10:     LOAD_COMMON2
yading@10:     unsigned map_generation = c->map_generation;
yading@10: 
yading@10:     av_assert1(minima_count <= MAX_SAB_SIZE);
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     /*Note j<MAX_SAB_SIZE is needed if MAX_SAB_SIZE < ME_MAP_SIZE as j can
yading@10:       become larger due to MVs overflowing their ME_MAP_MV_BITS bits space in map
yading@10:      */
yading@10:     for(j=i=0; i<ME_MAP_SIZE && j<MAX_SAB_SIZE; i++){
yading@10:         uint32_t key= map[i];
yading@10: 
yading@10:         key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));
yading@10: 
yading@10:         if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue;
yading@10: 
yading@10:         minima[j].height= score_map[i];
yading@10:         minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;
yading@10:         minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);
yading@10:         minima[j].x-= (1<<(ME_MAP_MV_BITS-1));
yading@10:         minima[j].y-= (1<<(ME_MAP_MV_BITS-1));
yading@10: 
yading@10:         // all entries in map should be in range except if the mv overflows their ME_MAP_MV_BITS bits space
yading@10:         if(   minima[j].x > xmax || minima[j].x < xmin
yading@10:            || minima[j].y > ymax || minima[j].y < ymin)
yading@10:             continue;
yading@10: 
yading@10:         minima[j].checked=0;
yading@10:         if(minima[j].x || minima[j].y)
yading@10:             minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;
yading@10: 
yading@10:         j++;
yading@10:     }
yading@10: 
yading@10:     qsort(minima, j, sizeof(Minima), minima_cmp);
yading@10: 
yading@10:     for(; j<minima_count; j++){
yading@10:         minima[j].height=256*256*256*64;
yading@10:         minima[j].checked=0;
yading@10:         minima[j].x= minima[j].y=0;
yading@10:     }
yading@10: 
yading@10:     for(i=0; i<minima_count; i++){
yading@10:         const int x= minima[i].x;
yading@10:         const int y= minima[i].y;
yading@10:         int d;
yading@10: 
yading@10:         if(minima[i].checked) continue;
yading@10: 
yading@10:         if(   x >= xmax || x <= xmin
yading@10:            || y >= ymax || y <= ymin)
yading@10:            continue;
yading@10: 
yading@10:         SAB_CHECK_MV(x-1, y)
yading@10:         SAB_CHECK_MV(x+1, y)
yading@10:         SAB_CHECK_MV(x  , y-1)
yading@10:         SAB_CHECK_MV(x  , y+1)
yading@10: 
yading@10:         minima[i].checked= 1;
yading@10:     }
yading@10: 
yading@10:     best[0]= minima[0].x;
yading@10:     best[1]= minima[0].y;
yading@10:     dmin= minima[0].height;
yading@10: 
yading@10:     if(   best[0] < xmax && best[0] > xmin
yading@10:        && best[1] < ymax && best[1] > ymin){
yading@10:         int d;
yading@10:         //ensure that the refernece samples for hpel refinement are in the map
yading@10:         CHECK_MV(best[0]-1, best[1])
yading@10:         CHECK_MV(best[0]+1, best[1])
yading@10:         CHECK_MV(best[0], best[1]-1)
yading@10:         CHECK_MV(best[0], best[1]+1)
yading@10:     }
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: static int var_diamond_search(MpegEncContext * s, int *best, int dmin,
yading@10:                                        int src_index, int ref_index, int const penalty_factor,
yading@10:                                        int size, int h, int flags)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     int dia_size;
yading@10:     LOAD_COMMON
yading@10:     LOAD_COMMON2
yading@10:     unsigned map_generation = c->map_generation;
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     for(dia_size=1; dia_size<=c->dia_size; dia_size++){
yading@10:         int dir, start, end;
yading@10:         const int x= best[0];
yading@10:         const int y= best[1];
yading@10: 
yading@10:         start= FFMAX(0, y + dia_size - ymax);
yading@10:         end  = FFMIN(dia_size, xmax - x + 1);
yading@10:         for(dir= start; dir<end; dir++){
yading@10:             int d;
yading@10: 
yading@10: //check(x + dir,y + dia_size - dir,0, a0)
yading@10:             CHECK_MV(x + dir           , y + dia_size - dir);
yading@10:         }
yading@10: 
yading@10:         start= FFMAX(0, x + dia_size - xmax);
yading@10:         end  = FFMIN(dia_size, y - ymin + 1);
yading@10:         for(dir= start; dir<end; dir++){
yading@10:             int d;
yading@10: 
yading@10: //check(x + dia_size - dir, y - dir,0, a1)
yading@10:             CHECK_MV(x + dia_size - dir, y - dir           );
yading@10:         }
yading@10: 
yading@10:         start= FFMAX(0, -y + dia_size + ymin );
yading@10:         end  = FFMIN(dia_size, x - xmin + 1);
yading@10:         for(dir= start; dir<end; dir++){
yading@10:             int d;
yading@10: 
yading@10: //check(x - dir,y - dia_size + dir,0, a2)
yading@10:             CHECK_MV(x - dir           , y - dia_size + dir);
yading@10:         }
yading@10: 
yading@10:         start= FFMAX(0, -x + dia_size + xmin );
yading@10:         end  = FFMIN(dia_size, ymax - y + 1);
yading@10:         for(dir= start; dir<end; dir++){
yading@10:             int d;
yading@10: 
yading@10: //check(x - dia_size + dir, y + dir,0, a3)
yading@10:             CHECK_MV(x - dia_size + dir, y + dir           );
yading@10:         }
yading@10: 
yading@10:         if(x!=best[0] || y!=best[1])
yading@10:             dia_size=0;
yading@10:     }
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: static av_always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
yading@10:                                        int src_index, int ref_index, int const penalty_factor,
yading@10:                                        int size, int h, int flags){
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     if(c->dia_size==-1)
yading@10:         return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10:     else if(c->dia_size<-1)
yading@10:         return   sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10:     else if(c->dia_size<2)
yading@10:         return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10:     else if(c->dia_size>1024)
yading@10:         return          full_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10:     else if(c->dia_size>768)
yading@10:         return           umh_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10:     else if(c->dia_size>512)
yading@10:         return           hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, c->dia_size&0xFF);
yading@10:     else if(c->dia_size>256)
yading@10:         return       l2s_dia_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10:     else
yading@10:         return   var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10: }
yading@10: 
yading@10: /**
yading@10:    @param P a list of candidate mvs to check before starting the
yading@10:    iterative search. If one of the candidates is close to the optimal mv, then
yading@10:    it takes fewer iterations. And it increases the chance that we find the
yading@10:    optimal mv.
yading@10:  */
yading@10: static av_always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
yading@10:                              int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
yading@10:                              int ref_mv_scale, int flags, int size, int h)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     int best[2]={0, 0};      /**< x and y coordinates of the best motion vector.
yading@10:                                i.e. the difference between the position of the
yading@10:                                block currently being encoded and the position of
yading@10:                                the block chosen to predict it from. */
yading@10:     int d;                   ///< the score (cmp + penalty) of any given mv
yading@10:     int dmin;                /**< the best value of d, i.e. the score
yading@10:                                corresponding to the mv stored in best[]. */
yading@10:     unsigned map_generation;
yading@10:     int penalty_factor;
yading@10:     const int ref_mv_stride= s->mb_stride; //pass as arg  FIXME
yading@10:     const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10: 
yading@10:     LOAD_COMMON
yading@10:     LOAD_COMMON2
yading@10: 
yading@10:     if(c->pre_pass){
yading@10:         penalty_factor= c->pre_penalty_factor;
yading@10:         cmpf= s->dsp.me_pre_cmp[size];
yading@10:         chroma_cmpf= s->dsp.me_pre_cmp[size+1];
yading@10:     }else{
yading@10:         penalty_factor= c->penalty_factor;
yading@10:         cmpf= s->dsp.me_cmp[size];
yading@10:         chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10:     }
yading@10: 
yading@10:     map_generation= update_map_generation(c);
yading@10: 
yading@10:     av_assert2(cmpf);
yading@10:     dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
yading@10:     map[0]= map_generation;
yading@10:     score_map[0]= dmin;
yading@10: 
yading@10:     //FIXME precalc first term below?
yading@10:     if((s->pict_type == AV_PICTURE_TYPE_B && !(c->flags & FLAG_DIRECT)) || s->flags&CODEC_FLAG_MV0)
yading@10:         dmin += (mv_penalty[pred_x] + mv_penalty[pred_y])*penalty_factor;
yading@10: 
yading@10:     /* first line */
yading@10:     if (s->first_slice_line) {
yading@10:         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
yading@10:         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
yading@10:     }else{
yading@10:         if(dmin<((h*h*s->avctx->mv0_threshold)>>8)
yading@10:                     && ( P_LEFT[0]    |P_LEFT[1]
yading@10:                         |P_TOP[0]     |P_TOP[1]
yading@10:                         |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
yading@10:             *mx_ptr= 0;
yading@10:             *my_ptr= 0;
yading@10:             c->skip=1;
yading@10:             return dmin;
yading@10:         }
yading@10:         CHECK_MV(    P_MEDIAN[0] >>shift ,    P_MEDIAN[1] >>shift)
yading@10:         CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)  , (P_MEDIAN[1]>>shift)-1)
yading@10:         CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)  , (P_MEDIAN[1]>>shift)+1)
yading@10:         CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)-1, (P_MEDIAN[1]>>shift)  )
yading@10:         CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)+1, (P_MEDIAN[1]>>shift)  )
yading@10:         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
yading@10:         CHECK_MV(P_LEFT[0]    >>shift, P_LEFT[1]    >>shift)
yading@10:         CHECK_MV(P_TOP[0]     >>shift, P_TOP[1]     >>shift)
yading@10:         CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
yading@10:     }
yading@10:     if(dmin>h*h*4){
yading@10:         if(c->pre_pass){
yading@10:             CHECK_CLIPPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                             (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
yading@10:             if(!s->first_slice_line)
yading@10:                 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                                 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
yading@10:         }else{
yading@10:             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                             (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
yading@10:             if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line
yading@10:                 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                                 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     if(c->avctx->last_predictor_count){
yading@10:         const int count= c->avctx->last_predictor_count;
yading@10:         const int xstart= FFMAX(0, s->mb_x - count);
yading@10:         const int ystart= FFMAX(0, s->mb_y - count);
yading@10:         const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
yading@10:         const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
yading@10:         int mb_y;
yading@10: 
yading@10:         for(mb_y=ystart; mb_y<yend; mb_y++){
yading@10:             int mb_x;
yading@10:             for(mb_x=xstart; mb_x<xend; mb_x++){
yading@10:                 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
yading@10:                 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
yading@10:                 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;
yading@10: 
yading@10:                 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
yading@10:                 CHECK_MV(mx,my)
yading@10:             }
yading@10:         }
yading@10:     }
yading@10: 
yading@10: //check(best[0],best[1],0, b0)
yading@10:     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10: 
yading@10: //check(best[0],best[1],0, b1)
yading@10:     *mx_ptr= best[0];
yading@10:     *my_ptr= best[1];
yading@10: 
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: //this function is dedicated to the braindamaged gcc
yading@10: int ff_epzs_motion_search(MpegEncContext *s, int *mx_ptr, int *my_ptr,
yading@10:                           int P[10][2], int src_index, int ref_index,
yading@10:                           int16_t (*last_mv)[2], int ref_mv_scale,
yading@10:                           int size, int h)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10: //FIXME convert other functions in the same way if faster
yading@10:     if(c->flags==0 && h==16 && size==0){
yading@10:         return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);
yading@10: //    case FLAG_QPEL:
yading@10: //        return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
yading@10:     }else{
yading@10:         return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);
yading@10:     }
yading@10: }
yading@10: 
yading@10: static int epzs_motion_search4(MpegEncContext * s,
yading@10:                              int *mx_ptr, int *my_ptr, int P[10][2],
yading@10:                              int src_index, int ref_index, int16_t (*last_mv)[2],
yading@10:                              int ref_mv_scale)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     int best[2]={0, 0};
yading@10:     int d, dmin;
yading@10:     unsigned map_generation;
yading@10:     const int penalty_factor= c->penalty_factor;
yading@10:     const int size=1;
yading@10:     const int h=8;
yading@10:     const int ref_mv_stride= s->mb_stride;
yading@10:     const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     LOAD_COMMON
yading@10:     int flags= c->flags;
yading@10:     LOAD_COMMON2
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     map_generation= update_map_generation(c);
yading@10: 
yading@10:     dmin = 1000000;
yading@10: 
yading@10:     /* first line */
yading@10:     if (s->first_slice_line) {
yading@10:         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
yading@10:         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
yading@10:         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
yading@10:     }else{
yading@10:         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
yading@10:         //FIXME try some early stop
yading@10:         CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
yading@10:         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
yading@10:         CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
yading@10:         CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
yading@10:         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
yading@10:     }
yading@10:     if(dmin>64*4){
yading@10:         CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                         (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
yading@10:         if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line
yading@10:             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                             (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
yading@10:     }
yading@10: 
yading@10:     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10: 
yading@10:     *mx_ptr= best[0];
yading@10:     *my_ptr= best[1];
yading@10: 
yading@10:     return dmin;
yading@10: }
yading@10: 
yading@10: //try to merge with above FIXME (needs PSNR test)
yading@10: static int epzs_motion_search2(MpegEncContext * s,
yading@10:                              int *mx_ptr, int *my_ptr, int P[10][2],
yading@10:                              int src_index, int ref_index, int16_t (*last_mv)[2],
yading@10:                              int ref_mv_scale)
yading@10: {
yading@10:     MotionEstContext * const c= &s->me;
yading@10:     int best[2]={0, 0};
yading@10:     int d, dmin;
yading@10:     unsigned map_generation;
yading@10:     const int penalty_factor= c->penalty_factor;
yading@10:     const int size=0; //FIXME pass as arg
yading@10:     const int h=8;
yading@10:     const int ref_mv_stride= s->mb_stride;
yading@10:     const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
yading@10:     me_cmp_func cmpf, chroma_cmpf;
yading@10:     LOAD_COMMON
yading@10:     int flags= c->flags;
yading@10:     LOAD_COMMON2
yading@10: 
yading@10:     cmpf= s->dsp.me_cmp[size];
yading@10:     chroma_cmpf= s->dsp.me_cmp[size+1];
yading@10: 
yading@10:     map_generation= update_map_generation(c);
yading@10: 
yading@10:     dmin = 1000000;
yading@10: 
yading@10:     /* first line */
yading@10:     if (s->first_slice_line) {
yading@10:         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
yading@10:         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
yading@10:         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
yading@10:     }else{
yading@10:         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
yading@10:         //FIXME try some early stop
yading@10:         CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
yading@10:         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
yading@10:         CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
yading@10:         CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
yading@10:         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
yading@10:     }
yading@10:     if(dmin>64*4){
yading@10:         CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                         (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
yading@10:         if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line
yading@10:             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
yading@10:                             (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
yading@10:     }
yading@10: 
yading@10:     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
yading@10: 
yading@10:     *mx_ptr= best[0];
yading@10:     *my_ptr= best[1];
yading@10: 
yading@10:     return dmin;
yading@10: }