falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
8114 Commits
32 Branches
511MB
Tree: edab3048c2
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'edab3048c2'
${ noResults }
FFmpeg/libavcodec/motion_est_template.c

1239 lines
43KB
Raw Blame History 

  1. /*

  2.  * Motion estimation

  3.  * Copyright (c) 2002-2004 Michael Niedermayer

  4.  *

  5.  * This file is part of FFmpeg.

  6.  *

  7.  * FFmpeg is free software; you can redistribute it and/or

  8.  * modify it under the terms of the GNU Lesser General Public

  9.  * License as published by the Free Software Foundation; either

  10.  * version 2.1 of the License, or (at your option) any later version.

  11.  *

  12.  * FFmpeg is distributed in the hope that it will be useful,

  13.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  14.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

  15.  * Lesser General Public License for more details.

  16.  *

  17.  * You should have received a copy of the GNU Lesser General Public

  18.  * License along with FFmpeg; if not, write to the Free Software

  19.  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

  20.  *

  21.  */

  22. 

  23. /**

  24.  * @file motion_est_template.c

  25.  * Motion estimation template.

  26.  */

  27. 

  28. //lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)

  29. #define LOAD_COMMON\

  30.     uint32_t attribute_unused * const score_map= c->score_map;\

  31.     const int attribute_unused xmin= c->xmin;\

  32.     const int attribute_unused ymin= c->ymin;\

  33.     const int attribute_unused xmax= c->xmax;\

  34.     const int attribute_unused ymax= c->ymax;\

  35.     uint8_t *mv_penalty= c->current_mv_penalty;\

  36.     const int pred_x= c->pred_x;\

  37.     const int pred_y= c->pred_y;\

  38. 

  39. #define CHECK_HALF_MV(dx, dy, x, y)\

  40. {\

  41.     const int hx= 2*(x)+(dx);\

  42.     const int hy= 2*(y)+(dy);\

  43.     d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\

  44.     d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\

  45.     COPY3_IF_LT(dmin, d, bx, hx, by, hy)\

  46. }

  47. 

  48. #if 0

  49. static int hpel_motion_search)(MpegEncContext * s,

  50.                                   int *mx_ptr, int *my_ptr, int dmin,

  51.                                   uint8_t *ref_data[3],

  52.                                   int size)

  53. {

  54.     const int xx = 16 * s->mb_x + 8*(n&1);

  55.     const int yy = 16 * s->mb_y + 8*(n>>1);

  56.     const int mx = *mx_ptr;

  57.     const int my = *my_ptr;

  58.     const int penalty_factor= c->sub_penalty_factor;

  59. 

  60.     LOAD_COMMON

  61. 

  62.  //   INIT;

  63.  //FIXME factorize

  64.     me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub;

  65. 

  66.     if(s->no_rounding /*FIXME b_type*/){

  67.         hpel_put= &s->dsp.put_no_rnd_pixels_tab[size];

  68.         chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1];

  69.     }else{

  70.         hpel_put=& s->dsp.put_pixels_tab[size];

  71.         chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];

  72.     }

  73.     cmpf= s->dsp.me_cmp[size];

  74.     chroma_cmpf= s->dsp.me_cmp[size+1];

  75.     cmp_sub= s->dsp.me_sub_cmp[size];

  76.     chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];

  77. 

  78.     if(c->skip){ //FIXME somehow move up (benchmark)

  79.         *mx_ptr = 0;

  80.         *my_ptr = 0;

  81.         return dmin;

  82.     }

  83. 

  84.     if(c->avctx->me_cmp != c->avctx->me_sub_cmp){

  85.         CMP_HPEL(dmin, 0, 0, mx, my, size);

  86.         if(mx || my)

  87.             dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;

  88.     }

  89. 

  90.     if (mx > xmin && mx < xmax &&

  91.         my > ymin && my < ymax) {

  92.         int bx=2*mx, by=2*my;

  93.         int d= dmin;

  94. 

  95.         CHECK_HALF_MV(1, 1, mx-1, my-1)

  96.         CHECK_HALF_MV(0, 1, mx  , my-1)

  97.         CHECK_HALF_MV(1, 1, mx  , my-1)

  98.         CHECK_HALF_MV(1, 0, mx-1, my  )

  99.         CHECK_HALF_MV(1, 0, mx  , my  )

  100.         CHECK_HALF_MV(1, 1, mx-1, my  )

  101.         CHECK_HALF_MV(0, 1, mx  , my  )

  102.         CHECK_HALF_MV(1, 1, mx  , my  )

  103. 

  104.         assert(bx >= xmin*2 || bx <= xmax*2 || by >= ymin*2 || by <= ymax*2);

  105. 

  106.         *mx_ptr = bx;

  107.         *my_ptr = by;

  108.     }else{

  109.         *mx_ptr =2*mx;

  110.         *my_ptr =2*my;

  111.     }

  112. 

  113.     return dmin;

  114. }

  115. 

  116. #else

  117. static int hpel_motion_search(MpegEncContext * s,

  118.                                   int *mx_ptr, int *my_ptr, int dmin,

  119.                                   int src_index, int ref_index,

  120.                                   int size, int h)

  121. {

  122.     MotionEstContext * const c= &s->me;

  123.     const int mx = *mx_ptr;

  124.     const int my = *my_ptr;

  125.     const int penalty_factor= c->sub_penalty_factor;

  126.     me_cmp_func cmp_sub, chroma_cmp_sub;

  127.     int bx=2*mx, by=2*my;

  128. 

  129.     LOAD_COMMON

  130.     int flags= c->sub_flags;

  131. 

  132.  //FIXME factorize

  133. 

  134.     cmp_sub= s->dsp.me_sub_cmp[size];

  135.     chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];

  136. 

  137.     if(c->skip){ //FIXME move out of hpel?

  138.         *mx_ptr = 0;

  139.         *my_ptr = 0;

  140.         return dmin;

  141.     }

  142. 

  143.     if(c->avctx->me_cmp != c->avctx->me_sub_cmp){

  144.         dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);

  145.         if(mx || my || size>0)

  146.             dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;

  147.     }

  148. 

  149.     if (mx > xmin && mx < xmax &&

  150.         my > ymin && my < ymax) {

  151.         int d= dmin;

  152.         const int index= (my<<ME_MAP_SHIFT) + mx;

  153.         const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]

  154.                      + (mv_penalty[bx   - pred_x] + mv_penalty[by-2 - pred_y])*c->penalty_factor;

  155.         const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)]

  156.                      + (mv_penalty[bx-2 - pred_x] + mv_penalty[by   - pred_y])*c->penalty_factor;

  157.         const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)]

  158.                      + (mv_penalty[bx+2 - pred_x] + mv_penalty[by   - pred_y])*c->penalty_factor;

  159.         const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]

  160.                      + (mv_penalty[bx   - pred_x] + mv_penalty[by+2 - pred_y])*c->penalty_factor;

  161. 

  162. #if 1

  163.         int key;

  164.         int map_generation= c->map_generation;

  165. #ifndef NDEBUG

  166.         uint32_t *map= c->map;

  167. #endif

  168.         key= ((my-1)<<ME_MAP_MV_BITS) + (mx) + map_generation;

  169.         assert(map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);

  170.         key= ((my+1)<<ME_MAP_MV_BITS) + (mx) + map_generation;

  171.         assert(map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);

  172.         key= ((my)<<ME_MAP_MV_BITS) + (mx+1) + map_generation;

  173.         assert(map[(index+1)&(ME_MAP_SIZE-1)] == key);

  174.         key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation;

  175.         assert(map[(index-1)&(ME_MAP_SIZE-1)] == key);

  176. #endif

  177.         if(t<=b){

  178.             CHECK_HALF_MV(0, 1, mx  ,my-1)

  179.             if(l<=r){

  180.                 CHECK_HALF_MV(1, 1, mx-1, my-1)

  181.                 if(t+r<=b+l){

  182.                     CHECK_HALF_MV(1, 1, mx  , my-1)

  183.                 }else{

  184.                     CHECK_HALF_MV(1, 1, mx-1, my  )

  185.                 }

  186.                 CHECK_HALF_MV(1, 0, mx-1, my  )

  187.             }else{

  188.                 CHECK_HALF_MV(1, 1, mx  , my-1)

  189.                 if(t+l<=b+r){

  190.                     CHECK_HALF_MV(1, 1, mx-1, my-1)

  191.                 }else{

  192.                     CHECK_HALF_MV(1, 1, mx  , my  )

  193.                 }

  194.                 CHECK_HALF_MV(1, 0, mx  , my  )

  195.             }

  196.         }else{

  197.             if(l<=r){

  198.                 if(t+l<=b+r){

  199.                     CHECK_HALF_MV(1, 1, mx-1, my-1)

  200.                 }else{

  201.                     CHECK_HALF_MV(1, 1, mx  , my  )

  202.                 }

  203.                 CHECK_HALF_MV(1, 0, mx-1, my)

  204.                 CHECK_HALF_MV(1, 1, mx-1, my)

  205.             }else{

  206.                 if(t+r<=b+l){

  207.                     CHECK_HALF_MV(1, 1, mx  , my-1)

  208.                 }else{

  209.                     CHECK_HALF_MV(1, 1, mx-1, my)

  210.                 }

  211.                 CHECK_HALF_MV(1, 0, mx  , my)

  212.                 CHECK_HALF_MV(1, 1, mx  , my)

  213.             }

  214.             CHECK_HALF_MV(0, 1, mx  , my)

  215.         }

  216.         assert(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2);

  217.     }

  218. 

  219.     *mx_ptr = bx;

  220.     *my_ptr = by;

  221. 

  222.     return dmin;

  223. }

  224. #endif

  225. 

  226. static int no_sub_motion_search(MpegEncContext * s,

  227.           int *mx_ptr, int *my_ptr, int dmin,

  228.                                   int src_index, int ref_index,

  229.                                   int size, int h)

  230. {

  231.     (*mx_ptr)<<=1;

  232.     (*my_ptr)<<=1;

  233.     return dmin;

  234. }

  235. 

  236. inline int ff_get_mb_score(MpegEncContext * s, int mx, int my, int src_index,

  237.                                int ref_index, int size, int h, int add_rate)

  238. {

  239. //    const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp;

  240.     MotionEstContext * const c= &s->me;

  241.     const int penalty_factor= c->mb_penalty_factor;

  242.     const int flags= c->mb_flags;

  243.     const int qpel= flags & FLAG_QPEL;

  244.     const int mask= 1+2*qpel;

  245.     me_cmp_func cmp_sub, chroma_cmp_sub;

  246.     int d;

  247. 

  248.     LOAD_COMMON

  249. 

  250.  //FIXME factorize

  251. 

  252.     cmp_sub= s->dsp.mb_cmp[size];

  253.     chroma_cmp_sub= s->dsp.mb_cmp[size+1];

  254. 

  255. //    assert(!c->skip);

  256. //    assert(c->avctx->me_sub_cmp != c->avctx->mb_cmp);

  257. 

  258.     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);

  259.     //FIXME check cbp before adding penalty for (0,0) vector

  260.     if(add_rate && (mx || my || size>0))

  261.         d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;

  262. 

  263.     return d;

  264. }

  265. 

  266. #define CHECK_QUARTER_MV(dx, dy, x, y)\

  267. {\

  268.     const int hx= 4*(x)+(dx);\

  269.     const int hy= 4*(y)+(dy);\

  270.     d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\

  271.     d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\

  272.     COPY3_IF_LT(dmin, d, bx, hx, by, hy)\

  273. }

  274. 

  275. static int qpel_motion_search(MpegEncContext * s,

  276.                                   int *mx_ptr, int *my_ptr, int dmin,

  277.                                   int src_index, int ref_index,

  278.                                   int size, int h)

  279. {

  280.     MotionEstContext * const c= &s->me;

  281.     const int mx = *mx_ptr;

  282.     const int my = *my_ptr;

  283.     const int penalty_factor= c->sub_penalty_factor;

  284.     const int map_generation= c->map_generation;

  285.     const int subpel_quality= c->avctx->me_subpel_quality;

  286.     uint32_t *map= c->map;

  287.     me_cmp_func cmpf, chroma_cmpf;

  288.     me_cmp_func cmp_sub, chroma_cmp_sub;

  289. 

  290.     LOAD_COMMON

  291.     int flags= c->sub_flags;

  292. 

  293.     cmpf= s->dsp.me_cmp[size];

  294.     chroma_cmpf= s->dsp.me_cmp[size+1]; //factorize FIXME

  295.  //FIXME factorize

  296. 

  297.     cmp_sub= s->dsp.me_sub_cmp[size];

  298.     chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];

  299. 

  300.     if(c->skip){ //FIXME somehow move up (benchmark)

  301.         *mx_ptr = 0;

  302.         *my_ptr = 0;

  303.         return dmin;

  304.     }

  305. 

  306.     if(c->avctx->me_cmp != c->avctx->me_sub_cmp){

  307.         dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);

  308.         if(mx || my || size>0)

  309.             dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor;

  310.     }

  311. 

  312.     if (mx > xmin && mx < xmax &&

  313.         my > ymin && my < ymax) {

  314.         int bx=4*mx, by=4*my;

  315.         int d= dmin;

  316.         int i, nx, ny;

  317.         const int index= (my<<ME_MAP_SHIFT) + mx;

  318.         const int t= score_map[(index-(1<<ME_MAP_SHIFT)  )&(ME_MAP_SIZE-1)];

  319.         const int l= score_map[(index- 1                 )&(ME_MAP_SIZE-1)];

  320.         const int r= score_map[(index+ 1                 )&(ME_MAP_SIZE-1)];

  321.         const int b= score_map[(index+(1<<ME_MAP_SHIFT)  )&(ME_MAP_SIZE-1)];

  322.         const int c= score_map[(index                    )&(ME_MAP_SIZE-1)];

  323.         int best[8];

  324.         int best_pos[8][2];

  325. 

  326.         memset(best, 64, sizeof(int)*8);

  327. #if 1

  328.         if(s->me.dia_size>=2){

  329.             const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];

  330.             const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];

  331.             const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];

  332.             const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];

  333. 

  334.             for(ny= -3; ny <= 3; ny++){

  335.                 for(nx= -3; nx <= 3; nx++){

  336.                     //FIXME this could overflow (unlikely though)

  337.                     const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;

  338.                     const int64_t c2= nx*nx*( r +  l - 2*c) + 4*nx*( r- l) + 32*c;

  339.                     const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b;

  340.                     int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10;

  341.                     int i;

  342. 

  343.                     if((nx&3)==0 && (ny&3)==0) continue;

  344. 

  345.                     score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;

  346. 

  347. //                    if(nx&1) score-=1024*c->penalty_factor;

  348. //                    if(ny&1) score-=1024*c->penalty_factor;

  349. 

  350.                     for(i=0; i<8; i++){

  351.                         if(score < best[i]){

  352.                             memmove(&best[i+1], &best[i], sizeof(int)*(7-i));

  353.                             memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));

  354.                             best[i]= score;

  355.                             best_pos[i][0]= nx + 4*mx;

  356.                             best_pos[i][1]= ny + 4*my;

  357.                             break;

  358.                         }

  359.                     }

  360.                 }

  361.             }

  362.         }else{

  363.             int tl;

  364.             //FIXME this could overflow (unlikely though)

  365.             const int cx = 4*(r - l);

  366.             const int cx2= r + l - 2*c;

  367.             const int cy = 4*(b - t);

  368.             const int cy2= b + t - 2*c;

  369.             int cxy;

  370. 

  371.             if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME

  372.                 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];

  373.             }else{

  374.                 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

  375.             }

  376. 

  377.             cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;

  378. 

  379.             assert(16*cx2 + 4*cx + 32*c == 32*r);

  380.             assert(16*cx2 - 4*cx + 32*c == 32*l);

  381.             assert(16*cy2 + 4*cy + 32*c == 32*b);

  382.             assert(16*cy2 - 4*cy + 32*c == 32*t);

  383.             assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);

  384. 

  385.             for(ny= -3; ny <= 3; ny++){

  386.                 for(nx= -3; nx <= 3; nx++){

  387.                     //FIXME this could overflow (unlikely though)

  388.                     int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor

  389.                     int i;

  390. 

  391.                     if((nx&3)==0 && (ny&3)==0) continue;

  392. 

  393.                     score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;

  394. //                    if(nx&1) score-=32*c->penalty_factor;

  395.   //                  if(ny&1) score-=32*c->penalty_factor;

  396. 

  397.                     for(i=0; i<8; i++){

  398.                         if(score < best[i]){

  399.                             memmove(&best[i+1], &best[i], sizeof(int)*(7-i));

  400.                             memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));

  401.                             best[i]= score;

  402.                             best_pos[i][0]= nx + 4*mx;

  403.                             best_pos[i][1]= ny + 4*my;

  404.                             break;

  405.                         }

  406.                     }

  407.                 }

  408.             }

  409.         }

  410.         for(i=0; i<subpel_quality; i++){

  411.             nx= best_pos[i][0];

  412.             ny= best_pos[i][1];

  413.             CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)

  414.         }

  415. 

  416. #if 0

  417.             const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];

  418.             const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];

  419.             const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];

  420.             const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];

  421. //            if(l < r && l < t && l < b && l < tl && l < bl && l < tr && l < br && bl < tl){

  422.             if(tl<br){

  423. 

  424. //            nx= FFMAX(4*mx - bx, bx - 4*mx);

  425. //            ny= FFMAX(4*my - by, by - 4*my);

  426. 

  427.             static int stats[7][7], count;

  428.             count++;

  429.             stats[4*mx - bx + 3][4*my - by + 3]++;

  430.             if(256*256*256*64 % count ==0){

  431.                 for(i=0; i<49; i++){

  432.                     if((i%7)==0) printf("\n");

  433.                     printf("%6d ", stats[0][i]);

  434.                 }

  435.                 printf("\n");

  436.             }

  437.             }

  438. #endif

  439. #else

  440. 

  441.         CHECK_QUARTER_MV(2, 2, mx-1, my-1)

  442.         CHECK_QUARTER_MV(0, 2, mx  , my-1)

  443.         CHECK_QUARTER_MV(2, 2, mx  , my-1)

  444.         CHECK_QUARTER_MV(2, 0, mx  , my  )

  445.         CHECK_QUARTER_MV(2, 2, mx  , my  )

  446.         CHECK_QUARTER_MV(0, 2, mx  , my  )

  447.         CHECK_QUARTER_MV(2, 2, mx-1, my  )

  448.         CHECK_QUARTER_MV(2, 0, mx-1, my  )

  449. 

  450.         nx= bx;

  451.         ny= by;

  452. 

  453.         for(i=0; i<8; i++){

  454.             int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1};

  455.             int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1};

  456.             CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>2, (ny + oy[i])>>2)

  457.         }

  458. #endif

  459. #if 0

  460.         //outer ring

  461.         CHECK_QUARTER_MV(1, 3, mx-1, my-1)

  462.         CHECK_QUARTER_MV(1, 2, mx-1, my-1)

  463.         CHECK_QUARTER_MV(1, 1, mx-1, my-1)

  464.         CHECK_QUARTER_MV(2, 1, mx-1, my-1)

  465.         CHECK_QUARTER_MV(3, 1, mx-1, my-1)

  466.         CHECK_QUARTER_MV(0, 1, mx  , my-1)

  467.         CHECK_QUARTER_MV(1, 1, mx  , my-1)

  468.         CHECK_QUARTER_MV(2, 1, mx  , my-1)

  469.         CHECK_QUARTER_MV(3, 1, mx  , my-1)

  470.         CHECK_QUARTER_MV(3, 2, mx  , my-1)

  471.         CHECK_QUARTER_MV(3, 3, mx  , my-1)

  472.         CHECK_QUARTER_MV(3, 0, mx  , my  )

  473.         CHECK_QUARTER_MV(3, 1, mx  , my  )

  474.         CHECK_QUARTER_MV(3, 2, mx  , my  )

  475.         CHECK_QUARTER_MV(3, 3, mx  , my  )

  476.         CHECK_QUARTER_MV(2, 3, mx  , my  )

  477.         CHECK_QUARTER_MV(1, 3, mx  , my  )

  478.         CHECK_QUARTER_MV(0, 3, mx  , my  )

  479.         CHECK_QUARTER_MV(3, 3, mx-1, my  )

  480.         CHECK_QUARTER_MV(2, 3, mx-1, my  )

  481.         CHECK_QUARTER_MV(1, 3, mx-1, my  )

  482.         CHECK_QUARTER_MV(1, 2, mx-1, my  )

  483.         CHECK_QUARTER_MV(1, 1, mx-1, my  )

  484.         CHECK_QUARTER_MV(1, 0, mx-1, my  )

  485. #endif

  486.         assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);

  487. 

  488.         *mx_ptr = bx;

  489.         *my_ptr = by;

  490.     }else{

  491.         *mx_ptr =4*mx;

  492.         *my_ptr =4*my;

  493.     }

  494. 

  495.     return dmin;

  496. }

  497. 

  498. 

  499. #define CHECK_MV(x,y)\

  500. {\

  501.     const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\

  502.     const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\

  503.     assert((x) >= xmin);\

  504.     assert((x) <= xmax);\

  505.     assert((y) >= ymin);\

  506.     assert((y) <= ymax);\

  507. /*printf("check_mv %d %d\n", x, y);*/\

  508.     if(map[index]!=key){\

  509.         d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\

  510.         map[index]= key;\

  511.         score_map[index]= d;\

  512.         d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\

  513. /*printf("score:%d\n", d);*/\

  514.         COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\

  515.     }\

  516. }

  517. 

  518. #define CHECK_CLIPPED_MV(ax,ay)\

  519. {\

  520.     const int Lx= ax;\

  521.     const int Ly= ay;\

  522.     const int Lx2= FFMAX(xmin, FFMIN(Lx, xmax));\

  523.     const int Ly2= FFMAX(ymin, FFMIN(Ly, ymax));\

  524.     CHECK_MV(Lx2, Ly2)\

  525. }

  526. 

  527. #define CHECK_MV_DIR(x,y,new_dir)\

  528. {\

  529.     const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\

  530.     const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\

  531. /*printf("check_mv_dir %d %d %d\n", x, y, new_dir);*/\

  532.     if(map[index]!=key){\

  533.         d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\

  534.         map[index]= key;\

  535.         score_map[index]= d;\

  536.         d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\

  537. /*printf("score:%d\n", d);*/\

  538.         if(d<dmin){\

  539.             best[0]=x;\

  540.             best[1]=y;\

  541.             dmin=d;\

  542.             next_dir= new_dir;\

  543.         }\

  544.     }\

  545. }

  546. 

  547. #define check(x,y,S,v)\

  548. if( (x)<(xmin<<(S)) ) printf("%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\

  549. if( (x)>(xmax<<(S)) ) printf("%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\

  550. if( (y)<(ymin<<(S)) ) printf("%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\

  551. if( (y)>(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\

  552. 

  553. #define LOAD_COMMON2\

  554.     uint32_t *map= c->map;\

  555.     const int qpel= flags&FLAG_QPEL;\

  556.     const int shift= 1+qpel;\

  557. 

  558. static av_always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,

  559.                                        int src_index, int ref_index, int const penalty_factor,

  560.                                        int size, int h, int flags)

  561. {

  562.     MotionEstContext * const c= &s->me;

  563.     me_cmp_func cmpf, chroma_cmpf;

  564.     int next_dir=-1;

  565.     LOAD_COMMON

  566.     LOAD_COMMON2

  567.     int map_generation= c->map_generation;

  568. 

  569.     cmpf= s->dsp.me_cmp[size];

  570.     chroma_cmpf= s->dsp.me_cmp[size+1];

  571. 

  572.     { /* ensure that the best point is in the MAP as h/qpel refinement needs it */

  573.         const int key= (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;

  574.         const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);

  575.         if(map[index]!=key){ //this will be executed only very rarey

  576.             score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);

  577.             map[index]= key;

  578.         }

  579.     }

  580. 

  581.     for(;;){

  582.         int d;

  583.         const int dir= next_dir;

  584.         const int x= best[0];

  585.         const int y= best[1];

  586.         next_dir=-1;

  587. 

  588. //printf("%d", dir);

  589.         if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y  , 0)

  590.         if(dir!=3 && y>ymin) CHECK_MV_DIR(x  , y-1, 1)

  591.         if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y  , 2)

  592.         if(dir!=1 && y<ymax) CHECK_MV_DIR(x  , y+1, 3)

  593. 

  594.         if(next_dir==-1){

  595.             return dmin;

  596.         }

  597.     }

  598. }

  599. 

  600. static int funny_diamond_search(MpegEncContext * s, int *best, int dmin,

  601.                                        int src_index, int ref_index, int const penalty_factor,

  602.                                        int size, int h, int flags)

  603. {

  604.     MotionEstContext * const c= &s->me;

  605.     me_cmp_func cmpf, chroma_cmpf;

  606.     int dia_size;

  607.     LOAD_COMMON

  608.     LOAD_COMMON2

  609.     int map_generation= c->map_generation;

  610. 

  611.     cmpf= s->dsp.me_cmp[size];

  612.     chroma_cmpf= s->dsp.me_cmp[size+1];

  613. 

  614.     for(dia_size=1; dia_size<=4; dia_size++){

  615.         int dir;

  616.         const int x= best[0];

  617.         const int y= best[1];

  618. 

  619.         if(dia_size&(dia_size-1)) continue;

  620. 

  621.         if(   x + dia_size > xmax

  622.            || x - dia_size < xmin

  623.            || y + dia_size > ymax

  624.            || y - dia_size < ymin)

  625.            continue;

  626. 

  627.         for(dir= 0; dir<dia_size; dir+=2){

  628.             int d;

  629. 

  630.             CHECK_MV(x + dir           , y + dia_size - dir);

  631.             CHECK_MV(x + dia_size - dir, y - dir           );

  632.             CHECK_MV(x - dir           , y - dia_size + dir);

  633.             CHECK_MV(x - dia_size + dir, y + dir           );

  634.         }

  635. 

  636.         if(x!=best[0] || y!=best[1])

  637.             dia_size=0;

  638. #if 0

  639. {

  640. int dx, dy, i;

  641. static int stats[8*8];

  642. dx= FFABS(x-best[0]);

  643. dy= FFABS(y-best[1]);

  644. if(dy>dx){

  645.     dx^=dy; dy^=dx; dx^=dy;

  646. }

  647. stats[dy*8 + dx] ++;

  648. if(256*256*256*64 % (stats[0]+1)==0){

  649.     for(i=0; i<64; i++){

  650.         if((i&7)==0) printf("\n");

  651.         printf("%8d ", stats[i]);

  652.     }

  653.     printf("\n");

  654. }

  655. }

  656. #endif

  657.     }

  658.     return dmin;

  659. }

  660. 

  661. static int hex_search(MpegEncContext * s, int *best, int dmin,

  662.                                        int src_index, int ref_index, int const penalty_factor,

  663.                                        int size, int h, int flags, int dia_size)

  664. {

  665.     MotionEstContext * const c= &s->me;

  666.     me_cmp_func cmpf, chroma_cmpf;

  667.     LOAD_COMMON

  668.     LOAD_COMMON2

  669.     int map_generation= c->map_generation;

  670.     int x,y,d;

  671.     const int dec= dia_size & (dia_size-1);

  672. 

  673.     cmpf= s->dsp.me_cmp[size];

  674.     chroma_cmpf= s->dsp.me_cmp[size+1];

  675. 

  676.     for(;dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){

  677.         do{

  678.             x= best[0];

  679.             y= best[1];

  680. 

  681.             CHECK_CLIPPED_MV(x  -dia_size    , y);

  682.             CHECK_CLIPPED_MV(x+  dia_size    , y);

  683.             CHECK_CLIPPED_MV(x+( dia_size>>1), y+dia_size);

  684.             CHECK_CLIPPED_MV(x+( dia_size>>1), y-dia_size);

  685.             if(dia_size>1){

  686.                 CHECK_CLIPPED_MV(x+(-dia_size>>1), y+dia_size);

  687.                 CHECK_CLIPPED_MV(x+(-dia_size>>1), y-dia_size);

  688.             }

  689.         }while(best[0] != x || best[1] != y);

  690.     }

  691. 

  692.     return dmin;

  693. }

  694. 

  695. static int l2s_dia_search(MpegEncContext * s, int *best, int dmin,

  696.                                        int src_index, int ref_index, int const penalty_factor,

  697.                                        int size, int h, int flags)

  698. {

  699.     MotionEstContext * const c= &s->me;

  700.     me_cmp_func cmpf, chroma_cmpf;

  701.     LOAD_COMMON

  702.     LOAD_COMMON2

  703.     int map_generation= c->map_generation;

  704.     int x,y,i,d;

  705.     int dia_size= c->dia_size&0xFF;

  706.     const int dec= dia_size & (dia_size-1);

  707.     static const int hex[8][2]={{-2, 0}, {-1,-1}, { 0,-2}, { 1,-1},

  708.                                 { 2, 0}, { 1, 1}, { 0, 2}, {-1, 1}};

  709. 

  710.     cmpf= s->dsp.me_cmp[size];

  711.     chroma_cmpf= s->dsp.me_cmp[size+1];

  712. 

  713.     for(; dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){

  714.         do{

  715.             x= best[0];

  716.             y= best[1];

  717.             for(i=0; i<8; i++){

  718.                 CHECK_CLIPPED_MV(x+hex[i][0]*dia_size, y+hex[i][1]*dia_size);

  719.             }

  720.         }while(best[0] != x || best[1] != y);

  721.     }

  722. 

  723.     x= best[0];

  724.     y= best[1];

  725.     CHECK_CLIPPED_MV(x+1, y);

  726.     CHECK_CLIPPED_MV(x, y+1);

  727.     CHECK_CLIPPED_MV(x-1, y);

  728.     CHECK_CLIPPED_MV(x, y-1);

  729. 

  730.     return dmin;

  731. }

  732. 

  733. static int umh_search(MpegEncContext * s, int *best, int dmin,

  734.                                        int src_index, int ref_index, int const penalty_factor,

  735.                                        int size, int h, int flags)

  736. {

  737.     MotionEstContext * const c= &s->me;

  738.     me_cmp_func cmpf, chroma_cmpf;

  739.     LOAD_COMMON

  740.     LOAD_COMMON2

  741.     int map_generation= c->map_generation;

  742.     int x,y,x2,y2, i, j, d;

  743.     const int dia_size= c->dia_size&0xFE;

  744.     static const int hex[16][2]={{-4,-2}, {-4,-1}, {-4, 0}, {-4, 1}, {-4, 2},

  745.                                  { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2},

  746.                                  {-2, 3}, { 0, 4}, { 2, 3},

  747.                                  {-2,-3}, { 0,-4}, { 2,-3},};

  748. 

  749.     cmpf= s->dsp.me_cmp[size];

  750.     chroma_cmpf= s->dsp.me_cmp[size+1];

  751. 

  752.     x= best[0];

  753.     y= best[1];

  754.     for(x2=FFMAX(x-dia_size+1, xmin); x2<=FFMIN(x+dia_size-1,xmax); x2+=2){

  755.         CHECK_MV(x2, y);

  756.     }

  757.     for(y2=FFMAX(y-dia_size/2+1, ymin); y2<=FFMIN(y+dia_size/2-1,ymax); y2+=2){

  758.         CHECK_MV(x, y2);

  759.     }

  760. 

  761.     x= best[0];

  762.     y= best[1];

  763.     for(y2=FFMAX(y-2, ymin); y2<=FFMIN(y+2,ymax); y2++){

  764.         for(x2=FFMAX(x-2, xmin); x2<=FFMIN(x+2,xmax); x2++){

  765.             CHECK_MV(x2, y2);

  766.         }

  767.     }

  768. 

  769. //FIXME prevent the CLIP stuff

  770. 

  771.     for(j=1; j<=dia_size/4; j++){

  772.         for(i=0; i<16; i++){

  773.             CHECK_CLIPPED_MV(x+hex[i][0]*j, y+hex[i][1]*j);

  774.         }

  775.     }

  776. 

  777.     return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, 2);

  778. }

  779. 

  780. #define SAB_CHECK_MV(ax,ay)\

  781. {\

  782.     const int key= ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\

  783.     const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\

  784. /*printf("sab check %d %d\n", ax, ay);*/\

  785.     if(map[index]!=key){\

  786.         d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\

  787.         map[index]= key;\

  788.         score_map[index]= d;\

  789.         d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\

  790. /*printf("score: %d\n", d);*/\

  791.         if(d < minima[minima_count-1].height){\

  792.             int j=0;\

  793.             \

  794.             while(d >= minima[j].height) j++;\

  795. \

  796.             memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\

  797. \

  798.             minima[j].checked= 0;\

  799.             minima[j].height= d;\

  800.             minima[j].x= ax;\

  801.             minima[j].y= ay;\

  802.             \

  803.             i=-1;\

  804.             continue;\

  805.         }\

  806.     }\

  807. }

  808. 

  809. #define MAX_SAB_SIZE ME_MAP_SIZE

  810. static int sab_diamond_search(MpegEncContext * s, int *best, int dmin,

  811.                                        int src_index, int ref_index, int const penalty_factor,

  812.                                        int size, int h, int flags)

  813. {

  814.     MotionEstContext * const c= &s->me;

  815.     me_cmp_func cmpf, chroma_cmpf;

  816.     Minima minima[MAX_SAB_SIZE];

  817.     const int minima_count= FFABS(c->dia_size);

  818.     int i, j;

  819.     LOAD_COMMON

  820.     LOAD_COMMON2

  821.     int map_generation= c->map_generation;

  822. 

  823.     cmpf= s->dsp.me_cmp[size];

  824.     chroma_cmpf= s->dsp.me_cmp[size+1];

  825. 

  826.     /*Note j<MAX_SAB_SIZE is needed if MAX_SAB_SIZE < ME_MAP_SIZE as j can

  827.       become larger due to MVs overflowing their ME_MAP_MV_BITS bits space in map

  828.      */

  829.     for(j=i=0; i<ME_MAP_SIZE && j<MAX_SAB_SIZE; i++){

  830.         uint32_t key= map[i];

  831. 

  832.         key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));

  833. 

  834.         if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue;

  835. 

  836.         minima[j].height= score_map[i];

  837.         minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;

  838.         minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);

  839.         minima[j].x-= (1<<(ME_MAP_MV_BITS-1));

  840.         minima[j].y-= (1<<(ME_MAP_MV_BITS-1));

  841. 

  842.         // all entries in map should be in range except if the mv overflows their ME_MAP_MV_BITS bits space

  843.         if(   minima[j].x > xmax || minima[j].x < xmin

  844.            || minima[j].y > ymax || minima[j].y < ymin)

  845.             continue;

  846. 

  847.         minima[j].checked=0;

  848.         if(minima[j].x || minima[j].y)

  849.             minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;

  850. 

  851.         j++;

  852.     }

  853. 

  854.     qsort(minima, j, sizeof(Minima), minima_cmp);

  855. 

  856.     for(; j<minima_count; j++){

  857.         minima[j].height=256*256*256*64;

  858.         minima[j].checked=0;

  859.         minima[j].x= minima[j].y=0;

  860.     }

  861. 

  862.     for(i=0; i<minima_count; i++){

  863.         const int x= minima[i].x;

  864.         const int y= minima[i].y;

  865.         int d;

  866. 

  867.         if(minima[i].checked) continue;

  868. 

  869.         if(   x >= xmax || x <= xmin

  870.            || y >= ymax || y <= ymin)

  871.            continue;

  872. 

  873.         SAB_CHECK_MV(x-1, y)

  874.         SAB_CHECK_MV(x+1, y)

  875.         SAB_CHECK_MV(x  , y-1)

  876.         SAB_CHECK_MV(x  , y+1)

  877. 

  878.         minima[i].checked= 1;

  879.     }

  880. 

  881.     best[0]= minima[0].x;

  882.     best[1]= minima[0].y;

  883.     dmin= minima[0].height;

  884. 

  885.     if(   best[0] < xmax && best[0] > xmin

  886.        && best[1] < ymax && best[1] > ymin){

  887.         int d;

  888.         //ensure that the refernece samples for hpel refinement are in the map

  889.         CHECK_MV(best[0]-1, best[1])

  890.         CHECK_MV(best[0]+1, best[1])

  891.         CHECK_MV(best[0], best[1]-1)

  892.         CHECK_MV(best[0], best[1]+1)

  893.     }

  894.     return dmin;

  895. }

  896. 

  897. static int var_diamond_search(MpegEncContext * s, int *best, int dmin,

  898.                                        int src_index, int ref_index, int const penalty_factor,

  899.                                        int size, int h, int flags)

  900. {

  901.     MotionEstContext * const c= &s->me;

  902.     me_cmp_func cmpf, chroma_cmpf;

  903.     int dia_size;

  904.     LOAD_COMMON

  905.     LOAD_COMMON2

  906.     int map_generation= c->map_generation;

  907. 

  908.     cmpf= s->dsp.me_cmp[size];

  909.     chroma_cmpf= s->dsp.me_cmp[size+1];

  910. 

  911.     for(dia_size=1; dia_size<=c->dia_size; dia_size++){

  912.         int dir, start, end;

  913.         const int x= best[0];

  914.         const int y= best[1];

  915. 

  916.         start= FFMAX(0, y + dia_size - ymax);

  917.         end  = FFMIN(dia_size, xmax - x + 1);

  918.         for(dir= start; dir<end; dir++){

  919.             int d;

  920. 

  921. //check(x + dir,y + dia_size - dir,0, a0)

  922.             CHECK_MV(x + dir           , y + dia_size - dir);

  923.         }

  924. 

  925.         start= FFMAX(0, x + dia_size - xmax);

  926.         end  = FFMIN(dia_size, y - ymin + 1);

  927.         for(dir= start; dir<end; dir++){

  928.             int d;

  929. 

  930. //check(x + dia_size - dir, y - dir,0, a1)

  931.             CHECK_MV(x + dia_size - dir, y - dir           );

  932.         }

  933. 

  934.         start= FFMAX(0, -y + dia_size + ymin );

  935.         end  = FFMIN(dia_size, x - xmin + 1);

  936.         for(dir= start; dir<end; dir++){

  937.             int d;

  938. 

  939. //check(x - dir,y - dia_size + dir,0, a2)

  940.             CHECK_MV(x - dir           , y - dia_size + dir);

  941.         }

  942. 

  943.         start= FFMAX(0, -x + dia_size + xmin );

  944.         end  = FFMIN(dia_size, ymax - y + 1);

  945.         for(dir= start; dir<end; dir++){

  946.             int d;

  947. 

  948. //check(x - dia_size + dir, y + dir,0, a3)

  949.             CHECK_MV(x - dia_size + dir, y + dir           );

  950.         }

  951. 

  952.         if(x!=best[0] || y!=best[1])

  953.             dia_size=0;

  954. #if 0

  955. {

  956. int dx, dy, i;

  957. static int stats[8*8];

  958. dx= FFABS(x-best[0]);

  959. dy= FFABS(y-best[1]);

  960. stats[dy*8 + dx] ++;

  961. if(256*256*256*64 % (stats[0]+1)==0){

  962.     for(i=0; i<64; i++){

  963.         if((i&7)==0) printf("\n");

  964.         printf("%6d ", stats[i]);

  965.     }

  966.     printf("\n");

  967. }

  968. }

  969. #endif

  970.     }

  971.     return dmin;

  972. }

  973. 

  974. static av_always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,

  975.                                        int src_index, int ref_index, int const penalty_factor,

  976.                                        int size, int h, int flags){

  977.     MotionEstContext * const c= &s->me;

  978.     if(c->dia_size==-1)

  979.         return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  980.     else if(c->dia_size<-1)

  981.         return   sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  982.     else if(c->dia_size<2)

  983.         return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  984.     else if(c->dia_size>768)

  985.         return           umh_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  986.     else if(c->dia_size>512)

  987.         return           hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, c->dia_size&0xFF);

  988.     else if(c->dia_size>256)

  989.         return       l2s_dia_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  990.     else

  991.         return   var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  992. }

  993. 

  994. static av_always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,

  995.                              int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],

  996.                              int ref_mv_scale, int flags, int size, int h)

  997. {

  998.     MotionEstContext * const c= &s->me;

  999.     int best[2]={0, 0};

  1000.     int d, dmin;

  1001.     int map_generation;

  1002.     int penalty_factor;

  1003.     const int ref_mv_stride= s->mb_stride; //pass as arg  FIXME

  1004.     const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME

  1005.     me_cmp_func cmpf, chroma_cmpf;

  1006. 

  1007.     LOAD_COMMON

  1008.     LOAD_COMMON2

  1009. 

  1010.     if(c->pre_pass){

  1011.         penalty_factor= c->pre_penalty_factor;

  1012.         cmpf= s->dsp.me_pre_cmp[size];

  1013.         chroma_cmpf= s->dsp.me_pre_cmp[size+1];

  1014.     }else{

  1015.         penalty_factor= c->penalty_factor;

  1016.         cmpf= s->dsp.me_cmp[size];

  1017.         chroma_cmpf= s->dsp.me_cmp[size+1];

  1018.     }

  1019. 

  1020.     map_generation= update_map_generation(c);

  1021. 

  1022.     assert(cmpf);

  1023.     dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);

  1024.     map[0]= map_generation;

  1025.     score_map[0]= dmin;

  1026. 

  1027.     //FIXME precalc first term below?

  1028.     if((s->pict_type == B_TYPE && !(c->flags & FLAG_DIRECT)) || s->flags&CODEC_FLAG_MV0)

  1029.         dmin += (mv_penalty[pred_x] + mv_penalty[pred_y])*penalty_factor;

  1030. 

  1031.     /* first line */

  1032.     if (s->first_slice_line) {

  1033.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1034.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1035.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1036.     }else{

  1037.         if(dmin<((h*h*s->avctx->mv0_threshold)>>8)

  1038.                     && ( P_LEFT[0]    |P_LEFT[1]

  1039.                         |P_TOP[0]     |P_TOP[1]

  1040.                         |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){

  1041.             *mx_ptr= 0;

  1042.             *my_ptr= 0;

  1043.             c->skip=1;

  1044.             return dmin;

  1045.         }

  1046.         CHECK_MV(    P_MEDIAN[0] >>shift ,    P_MEDIAN[1] >>shift)

  1047.         CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)  , (P_MEDIAN[1]>>shift)-1)

  1048.         CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)  , (P_MEDIAN[1]>>shift)+1)

  1049.         CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)-1, (P_MEDIAN[1]>>shift)  )

  1050.         CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)+1, (P_MEDIAN[1]>>shift)  )

  1051.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1052.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1053.         CHECK_MV(P_LEFT[0]    >>shift, P_LEFT[1]    >>shift)

  1054.         CHECK_MV(P_TOP[0]     >>shift, P_TOP[1]     >>shift)

  1055.         CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)

  1056.     }

  1057.     if(dmin>h*h*4){

  1058.         if(c->pre_pass){

  1059.             CHECK_CLIPPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,

  1060.                             (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)

  1061.             if(!s->first_slice_line)

  1062.                 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,

  1063.                                 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)

  1064.         }else{

  1065.             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,

  1066.                             (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)

  1067.             if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line

  1068.                 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,

  1069.                                 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)

  1070.         }

  1071.     }

  1072. 

  1073.     if(c->avctx->last_predictor_count){

  1074.         const int count= c->avctx->last_predictor_count;

  1075.         const int xstart= FFMAX(0, s->mb_x - count);

  1076.         const int ystart= FFMAX(0, s->mb_y - count);

  1077.         const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);

  1078.         const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);

  1079.         int mb_y;

  1080. 

  1081.         for(mb_y=ystart; mb_y<yend; mb_y++){

  1082.             int mb_x;

  1083.             for(mb_x=xstart; mb_x<xend; mb_x++){

  1084.                 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;

  1085.                 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;

  1086.                 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;

  1087. 

  1088.                 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;

  1089.                 CHECK_MV(mx,my)

  1090.             }

  1091.         }

  1092.     }

  1093. 

  1094. //check(best[0],best[1],0, b0)

  1095.     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  1096. 

  1097. //check(best[0],best[1],0, b1)

  1098.     *mx_ptr= best[0];

  1099.     *my_ptr= best[1];

  1100. 

  1101. //    printf("%d %d %d \n", best[0], best[1], dmin);

  1102.     return dmin;

  1103. }

  1104. 

  1105. //this function is dedicated to the braindamaged gcc

  1106. inline int ff_epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr,

  1107.                              int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],

  1108.                              int ref_mv_scale, int size, int h)

  1109. {

  1110.     MotionEstContext * const c= &s->me;

  1111. //FIXME convert other functions in the same way if faster

  1112.     if(c->flags==0 && h==16 && size==0){

  1113.         return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);

  1114. //    case FLAG_QPEL:

  1115. //        return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);

  1116.     }else{

  1117.         return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);

  1118.     }

  1119. }

  1120. 

  1121. static int epzs_motion_search4(MpegEncContext * s,

  1122.                              int *mx_ptr, int *my_ptr, int P[10][2],

  1123.                              int src_index, int ref_index, int16_t (*last_mv)[2],

  1124.                              int ref_mv_scale)

  1125. {

  1126.     MotionEstContext * const c= &s->me;

  1127.     int best[2]={0, 0};

  1128.     int d, dmin;

  1129.     int map_generation;

  1130.     const int penalty_factor= c->penalty_factor;

  1131.     const int size=1;

  1132.     const int h=8;

  1133.     const int ref_mv_stride= s->mb_stride;

  1134.     const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;

  1135.     me_cmp_func cmpf, chroma_cmpf;

  1136.     LOAD_COMMON

  1137.     int flags= c->flags;

  1138.     LOAD_COMMON2

  1139. 

  1140.     cmpf= s->dsp.me_cmp[size];

  1141.     chroma_cmpf= s->dsp.me_cmp[size+1];

  1142. 

  1143.     map_generation= update_map_generation(c);

  1144. 

  1145.     dmin = 1000000;

  1146. //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);

  1147.     /* first line */

  1148.     if (s->first_slice_line) {

  1149.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1150.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1151.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1152.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1153.     }else{

  1154.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1155.         //FIXME try some early stop

  1156.         CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)

  1157.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1158.         CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)

  1159.         CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)

  1160.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1161.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1162.     }

  1163.     if(dmin>64*4){

  1164.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,

  1165.                         (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)

  1166.         if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line

  1167.             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,

  1168.                             (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)

  1169.     }

  1170. 

  1171.     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  1172. 

  1173.     *mx_ptr= best[0];

  1174.     *my_ptr= best[1];

  1175. 

  1176. //    printf("%d %d %d \n", best[0], best[1], dmin);

  1177.     return dmin;

  1178. }

  1179. 

  1180. //try to merge with above FIXME (needs PSNR test)

  1181. static int epzs_motion_search2(MpegEncContext * s,

  1182.                              int *mx_ptr, int *my_ptr, int P[10][2],

  1183.                              int src_index, int ref_index, int16_t (*last_mv)[2],

  1184.                              int ref_mv_scale)

  1185. {

  1186.     MotionEstContext * const c= &s->me;

  1187.     int best[2]={0, 0};

  1188.     int d, dmin;

  1189.     int map_generation;

  1190.     const int penalty_factor= c->penalty_factor;

  1191.     const int size=0; //FIXME pass as arg

  1192.     const int h=8;

  1193.     const int ref_mv_stride= s->mb_stride;

  1194.     const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;

  1195.     me_cmp_func cmpf, chroma_cmpf;

  1196.     LOAD_COMMON

  1197.     int flags= c->flags;

  1198.     LOAD_COMMON2

  1199. 

  1200.     cmpf= s->dsp.me_cmp[size];

  1201.     chroma_cmpf= s->dsp.me_cmp[size+1];

  1202. 

  1203.     map_generation= update_map_generation(c);

  1204. 

  1205.     dmin = 1000000;

  1206. //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);

  1207.     /* first line */

  1208.     if (s->first_slice_line) {

  1209.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1210.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1211.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1212.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1213.     }else{

  1214.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1215.         //FIXME try some early stop

  1216.         CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)

  1217.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1218.         CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)

  1219.         CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)

  1220.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1221.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1222.     }

  1223.     if(dmin>64*4){

  1224.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,

  1225.                         (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)

  1226.         if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line

  1227.             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,

  1228.                             (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)

  1229.     }

  1230. 

  1231.     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  1232. 

  1233.     *mx_ptr= best[0];

  1234.     *my_ptr= best[1];

  1235. 

  1236. //    printf("%d %d %d \n", best[0], best[1], dmin);

  1237.     return dmin;

  1238. }