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.
44792 Commits
32 Branches
369MB
Tree: ff0c628268
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'ff0c628268'
${ noResults }
FFmpeg/libavcodec/motion_est_template.c

1112 lines
40KB
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.  * @file

  24.  * Motion estimation template.

  25.  */

  26. 

  27. //Let us hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)

  28. #define LOAD_COMMON\

  29.     uint32_t av_unused * const score_map= c->score_map;\

  30.     const int av_unused xmin= c->xmin;\

  31.     const int av_unused ymin= c->ymin;\

  32.     const int av_unused xmax= c->xmax;\

  33.     const int av_unused ymax= c->ymax;\

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

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

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

  37. 

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

  39. {\

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

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

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

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

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

  45. }

  46. 

  47. static int hpel_motion_search(MpegEncContext * s,

  48.                                   int *mx_ptr, int *my_ptr, int dmin,

  49.                                   int src_index, int ref_index,

  50.                                   int size, int h)

  51. {

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

  53.     const int mx = *mx_ptr;

  54.     const int my = *my_ptr;

  55.     const int penalty_factor= c->sub_penalty_factor;

  56.     me_cmp_func cmp_sub, chroma_cmp_sub;

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

  58. 

  59.     LOAD_COMMON

  60.     int flags= c->sub_flags;

  61. 

  62.  //FIXME factorize

  63. 

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

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

  66. 

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

  68.         *mx_ptr = 0;

  69.         *my_ptr = 0;

  70.         return dmin;

  71.     }

  72. 

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

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

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

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

  77.     }

  78. 

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

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

  81.         int d= dmin;

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

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

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

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

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

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

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

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

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

  91. 

  92. #if 1

  93.         unsigned key;

  94.         unsigned map_generation= c->map_generation;

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

  96.         av_assert2(c->map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);

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

  98.         av_assert2(c->map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);

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

  100.         av_assert2(c->map[(index+1)&(ME_MAP_SIZE-1)] == key);

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

  102.         av_assert2(c->map[(index-1)&(ME_MAP_SIZE-1)] == key);

  103. #endif

  104.         if(t<=b){

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

  106.             if(l<=r){

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

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

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

  110.                 }else{

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

  112.                 }

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

  114.             }else{

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

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

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

  118.                 }else{

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

  120.                 }

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

  122.             }

  123.         }else{

  124.             if(l<=r){

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

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

  127.                 }else{

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

  129.                 }

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

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

  132.             }else{

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

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

  135.                 }else{

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

  137.                 }

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

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

  140.             }

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

  142.         }

  143.         av_assert2(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2);

  144.     }

  145. 

  146.     *mx_ptr = bx;

  147.     *my_ptr = by;

  148. 

  149.     return dmin;

  150. }

  151. 

  152. static int no_sub_motion_search(MpegEncContext * s,

  153.           int *mx_ptr, int *my_ptr, int dmin,

  154.                                   int src_index, int ref_index,

  155.                                   int size, int h)

  156. {

  157.     (*mx_ptr)<<=1;

  158.     (*my_ptr)<<=1;

  159.     return dmin;

  160. }

  161. 

  162. static inline int get_mb_score(MpegEncContext *s, int mx, int my,

  163.                                int src_index, int ref_index, int size,

  164.                                int h, int add_rate)

  165. {

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

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

  168.     const int penalty_factor= c->mb_penalty_factor;

  169.     const int flags= c->mb_flags;

  170.     const int qpel= flags & FLAG_QPEL;

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

  172.     me_cmp_func cmp_sub, chroma_cmp_sub;

  173.     int d;

  174. 

  175.     LOAD_COMMON

  176. 

  177.  //FIXME factorize

  178. 

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

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

  181. 

  182.     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);

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

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

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

  186. 

  187.     return d;

  188. }

  189. 

  190. int ff_get_mb_score(MpegEncContext *s, int mx, int my, int src_index,

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

  192. {

  193.     return get_mb_score(s, mx, my, src_index, ref_index, size, h, add_rate);

  194. }

  195. 

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

  197. {\

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

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

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

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

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

  203. }

  204. 

  205. static int qpel_motion_search(MpegEncContext * s,

  206.                                   int *mx_ptr, int *my_ptr, int dmin,

  207.                                   int src_index, int ref_index,

  208.                                   int size, int h)

  209. {

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

  211.     const int mx = *mx_ptr;

  212.     const int my = *my_ptr;

  213.     const int penalty_factor= c->sub_penalty_factor;

  214.     const unsigned map_generation = c->map_generation;

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

  216.     uint32_t *map= c->map;

  217.     me_cmp_func cmpf, chroma_cmpf;

  218.     me_cmp_func cmp_sub, chroma_cmp_sub;

  219. 

  220.     LOAD_COMMON

  221.     int flags= c->sub_flags;

  222. 

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

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

  225.  //FIXME factorize

  226. 

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

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

  229. 

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

  231.         *mx_ptr = 0;

  232.         *my_ptr = 0;

  233.         return dmin;

  234.     }

  235. 

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

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

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

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

  240.     }

  241. 

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

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

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

  245.         int d= dmin;

  246.         int i, nx, ny;

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

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

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

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

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

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

  253.         int best[8];

  254.         int best_pos[8][2];

  255. 

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

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

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

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

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

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

  262. 

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

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

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

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

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

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

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

  270.                     int i;

  271. 

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

  273. 

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

  275. 

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

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

  278. 

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

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

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

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

  283.                             best[i]= score;

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

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

  286.                             break;

  287.                         }

  288.                     }

  289.                 }

  290.             }

  291.         }else{

  292.             int tl;

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

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

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

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

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

  298.             int cxy;

  299. 

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

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

  302.             }else{

  303.                 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

  304.             }

  305. 

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

  307. 

  308.             av_assert2(16*cx2 + 4*cx + 32*c == 32*r);

  309.             av_assert2(16*cx2 - 4*cx + 32*c == 32*l);

  310.             av_assert2(16*cy2 + 4*cy + 32*c == 32*b);

  311.             av_assert2(16*cy2 - 4*cy + 32*c == 32*t);

  312.             av_assert2(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);

  313. 

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

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

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

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

  318.                     int i;

  319. 

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

  321. 

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

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

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

  325. 

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

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

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

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

  330.                             best[i]= score;

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

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

  333.                             break;

  334.                         }

  335.                     }

  336.                 }

  337.             }

  338.         }

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

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

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

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

  343.         }

  344. 

  345.         av_assert2(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);

  346. 

  347.         *mx_ptr = bx;

  348.         *my_ptr = by;

  349.     }else{

  350.         *mx_ptr =4*mx;

  351.         *my_ptr =4*my;

  352.     }

  353. 

  354.     return dmin;

  355. }

  356. 

  357. 

  358. #define CHECK_MV(x,y)\

  359. {\

  360.     const unsigned key = ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\

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

  362.     av_assert2((x) >= xmin);\

  363.     av_assert2((x) <= xmax);\

  364.     av_assert2((y) >= ymin);\

  365.     av_assert2((y) <= ymax);\

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

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

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

  369.         map[index]= key;\

  370.         score_map[index]= d;\

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

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

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

  374.     }\

  375. }

  376. 

  377. #define CHECK_CLIPPED_MV(ax,ay)\

  378. {\

  379.     const int Lx= ax;\

  380.     const int Ly= ay;\

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

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

  383.     CHECK_MV(Lx2, Ly2)\

  384. }

  385. 

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

  387. {\

  388.     const unsigned key = ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\

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

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

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

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

  393.         map[index]= key;\

  394.         score_map[index]= d;\

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

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

  397.         if(d<dmin){\

  398.             best[0]=x;\

  399.             best[1]=y;\

  400.             dmin=d;\

  401.             next_dir= new_dir;\

  402.         }\

  403.     }\

  404. }

  405. 

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

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

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

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

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

  411. 

  412. #define LOAD_COMMON2\

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

  414.     const int qpel= flags&FLAG_QPEL;\

  415.     const int shift= 1+qpel;\

  416. 

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

  418.                                        int src_index, int ref_index, int const penalty_factor,

  419.                                        int size, int h, int flags)

  420. {

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

  422.     me_cmp_func cmpf, chroma_cmpf;

  423.     int next_dir=-1;

  424.     LOAD_COMMON

  425.     LOAD_COMMON2

  426.     unsigned map_generation = c->map_generation;

  427. 

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

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

  430. 

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

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

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

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

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

  436.             map[index]= key;

  437.         }

  438.     }

  439. 

  440.     for(;;){

  441.         int d;

  442.         const int dir= next_dir;

  443.         const int x= best[0];

  444.         const int y= best[1];

  445.         next_dir=-1;

  446. 

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

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

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

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

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

  452. 

  453.         if(next_dir==-1){

  454.             return dmin;

  455.         }

  456.     }

  457. }

  458. 

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

  460.                                        int src_index, int ref_index, int const penalty_factor,

  461.                                        int size, int h, int flags)

  462. {

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

  464.     me_cmp_func cmpf, chroma_cmpf;

  465.     int dia_size;

  466.     LOAD_COMMON

  467.     LOAD_COMMON2

  468.     unsigned map_generation = c->map_generation;

  469. 

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

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

  472. 

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

  474.         int dir;

  475.         const int x= best[0];

  476.         const int y= best[1];

  477. 

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

  479. 

  480.         if(   x + dia_size > xmax

  481.            || x - dia_size < xmin

  482.            || y + dia_size > ymax

  483.            || y - dia_size < ymin)

  484.            continue;

  485. 

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

  487.             int d;

  488. 

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

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

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

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

  493.         }

  494. 

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

  496.             dia_size=0;

  497.     }

  498.     return dmin;

  499. }

  500. 

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

  502.                                        int src_index, int ref_index, int const penalty_factor,

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

  504. {

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

  506.     me_cmp_func cmpf, chroma_cmpf;

  507.     LOAD_COMMON

  508.     LOAD_COMMON2

  509.     unsigned map_generation = c->map_generation;

  510.     int x,y,d;

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

  512. 

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

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

  515. 

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

  517.         do{

  518.             x= best[0];

  519.             y= best[1];

  520. 

  521.             CHECK_CLIPPED_MV(x  -dia_size    , y);

  522.             CHECK_CLIPPED_MV(x+  dia_size    , y);

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

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

  525.             if(dia_size>1){

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

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

  528.             }

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

  530.     }

  531. 

  532.     return dmin;

  533. }

  534. 

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

  536.                                        int src_index, int ref_index, int const penalty_factor,

  537.                                        int size, int h, int flags)

  538. {

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

  540.     me_cmp_func cmpf, chroma_cmpf;

  541.     LOAD_COMMON

  542.     LOAD_COMMON2

  543.     unsigned map_generation = c->map_generation;

  544.     int x,y,i,d;

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

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

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

  548.                                 { 2, 0}, { 1, 1}, { 0, 2}, {-1, 1}};

  549. 

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

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

  552. 

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

  554.         do{

  555.             x= best[0];

  556.             y= best[1];

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

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

  559.             }

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

  561.     }

  562. 

  563.     x= best[0];

  564.     y= best[1];

  565.     CHECK_CLIPPED_MV(x+1, y);

  566.     CHECK_CLIPPED_MV(x, y+1);

  567.     CHECK_CLIPPED_MV(x-1, y);

  568.     CHECK_CLIPPED_MV(x, y-1);

  569. 

  570.     return dmin;

  571. }

  572. 

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

  574.                                        int src_index, int ref_index, int const penalty_factor,

  575.                                        int size, int h, int flags)

  576. {

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

  578.     me_cmp_func cmpf, chroma_cmpf;

  579.     LOAD_COMMON

  580.     LOAD_COMMON2

  581.     unsigned map_generation = c->map_generation;

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

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

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

  585.                                  { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2},

  586.                                  {-2, 3}, { 0, 4}, { 2, 3},

  587.                                  {-2,-3}, { 0,-4}, { 2,-3},};

  588. 

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

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

  591. 

  592.     x= best[0];

  593.     y= best[1];

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

  595.         CHECK_MV(x2, y);

  596.     }

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

  598.         CHECK_MV(x, y2);

  599.     }

  600. 

  601.     x= best[0];

  602.     y= best[1];

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

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

  605.             CHECK_MV(x2, y2);

  606.         }

  607.     }

  608. 

  609. //FIXME prevent the CLIP stuff

  610. 

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

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

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

  614.         }

  615.     }

  616. 

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

  618. }

  619. 

  620. static int full_search(MpegEncContext * s, int *best, int dmin,

  621.                                        int src_index, int ref_index, int const penalty_factor,

  622.                                        int size, int h, int flags)

  623. {

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

  625.     me_cmp_func cmpf, chroma_cmpf;

  626.     LOAD_COMMON

  627.     LOAD_COMMON2

  628.     unsigned map_generation = c->map_generation;

  629.     int x,y, d;

  630.     const int dia_size= c->dia_size&0xFF;

  631. 

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

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

  634. 

  635.     for(y=FFMAX(-dia_size, ymin); y<=FFMIN(dia_size,ymax); y++){

  636.         for(x=FFMAX(-dia_size, xmin); x<=FFMIN(dia_size,xmax); x++){

  637.             CHECK_MV(x, y);

  638.         }

  639.     }

  640. 

  641.     x= best[0];

  642.     y= best[1];

  643.     d= dmin;

  644.     CHECK_CLIPPED_MV(x  , y);

  645.     CHECK_CLIPPED_MV(x+1, y);

  646.     CHECK_CLIPPED_MV(x, y+1);

  647.     CHECK_CLIPPED_MV(x-1, y);

  648.     CHECK_CLIPPED_MV(x, y-1);

  649.     best[0]= x;

  650.     best[1]= y;

  651. 

  652.     return d;

  653. }

  654. 

  655. #define SAB_CHECK_MV(ax,ay)\

  656. {\

  657.     const unsigned key = ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\

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

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

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

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

  662.         map[index]= key;\

  663.         score_map[index]= d;\

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

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

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

  667.             int j=0;\

  668.             \

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

  670. \

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

  672. \

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

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

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

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

  677.             \

  678.             i=-1;\

  679.             continue;\

  680.         }\

  681.     }\

  682. }

  683. 

  684. #define MAX_SAB_SIZE ME_MAP_SIZE

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

  686.                                        int src_index, int ref_index, int const penalty_factor,

  687.                                        int size, int h, int flags)

  688. {

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

  690.     me_cmp_func cmpf, chroma_cmpf;

  691.     Minima minima[MAX_SAB_SIZE];

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

  693.     int i, j;

  694.     LOAD_COMMON

  695.     LOAD_COMMON2

  696.     unsigned map_generation = c->map_generation;

  697. 

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

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

  700. 

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

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

  703.      */

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

  705.         uint32_t key= map[i];

  706. 

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

  708. 

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

  710. 

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

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

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

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

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

  716. 

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

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

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

  720.             continue;

  721. 

  722.         minima[j].checked=0;

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

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

  725. 

  726.         j++;

  727.     }

  728. 

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

  730. 

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

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

  733.         minima[j].checked=0;

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

  735.     }

  736. 

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

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

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

  740.         int d;

  741. 

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

  743. 

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

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

  746.            continue;

  747. 

  748.         SAB_CHECK_MV(x-1, y)

  749.         SAB_CHECK_MV(x+1, y)

  750.         SAB_CHECK_MV(x  , y-1)

  751.         SAB_CHECK_MV(x  , y+1)

  752. 

  753.         minima[i].checked= 1;

  754.     }

  755. 

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

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

  758.     dmin= minima[0].height;

  759. 

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

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

  762.         int d;

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

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

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

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

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

  768.     }

  769.     return dmin;

  770. }

  771. 

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

  773.                                        int src_index, int ref_index, int const penalty_factor,

  774.                                        int size, int h, int flags)

  775. {

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

  777.     me_cmp_func cmpf, chroma_cmpf;

  778.     int dia_size;

  779.     LOAD_COMMON

  780.     LOAD_COMMON2

  781.     unsigned map_generation = c->map_generation;

  782. 

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

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

  785. 

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

  787.         int dir, start, end;

  788.         const int x= best[0];

  789.         const int y= best[1];

  790. 

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

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

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

  794.             int d;

  795. 

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

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

  798.         }

  799. 

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

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

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

  803.             int d;

  804. 

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

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

  807.         }

  808. 

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

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

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

  812.             int d;

  813. 

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

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

  816.         }

  817. 

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

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

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

  821.             int d;

  822. 

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

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

  825.         }

  826. 

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

  828.             dia_size=0;

  829.     }

  830.     return dmin;

  831. }

  832. 

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

  834.                                        int src_index, int ref_index, int const penalty_factor,

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

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

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

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

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

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

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

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

  843.     else if(c->dia_size>1024)

  844.         return          full_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

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

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

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

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

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

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

  851.     else

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

  853. }

  854. 

  855. /**

  856.    @param P a list of candidate mvs to check before starting the

  857.    iterative search. If one of the candidates is close to the optimal mv, then

  858.    it takes fewer iterations. And it increases the chance that we find the

  859.    optimal mv.

  860.  */

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

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

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

  864. {

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

  866.     int best[2]={0, 0};      /**< x and y coordinates of the best motion vector.

  867.                                i.e. the difference between the position of the

  868.                                block currently being encoded and the position of

  869.                                the block chosen to predict it from. */

  870.     int d;                   ///< the score (cmp + penalty) of any given mv

  871.     int dmin;                /**< the best value of d, i.e. the score

  872.                                corresponding to the mv stored in best[]. */

  873.     unsigned map_generation;

  874.     int penalty_factor;

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

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

  877.     me_cmp_func cmpf, chroma_cmpf;

  878. 

  879.     LOAD_COMMON

  880.     LOAD_COMMON2

  881. 

  882.     if(c->pre_pass){

  883.         penalty_factor= c->pre_penalty_factor;

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

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

  886.     }else{

  887.         penalty_factor= c->penalty_factor;

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

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

  890.     }

  891. 

  892.     map_generation= update_map_generation(c);

  893. 

  894.     av_assert2(cmpf);

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

  896.     map[0]= map_generation;

  897.     score_map[0]= dmin;

  898. 

  899.     //FIXME precalc first term below?

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

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

  902. 

  903.     /* first line */

  904.     if (s->first_slice_line) {

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

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

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

  908.     }else{

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

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

  911.                         |P_TOP[0]     |P_TOP[1]

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

  913.             *mx_ptr= 0;

  914.             *my_ptr= 0;

  915.             c->skip=1;

  916.             return dmin;

  917.         }

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

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

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

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

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

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

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

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

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

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

  928.     }

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

  930.         if(c->pre_pass){

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

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

  933.             if(!s->first_slice_line)

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

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

  936.         }else{

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

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

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

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

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

  942.         }

  943.     }

  944. 

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

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

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

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

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

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

  951.         int mb_y;

  952. 

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

  954.             int mb_x;

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

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

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

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

  959. 

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

  961.                 CHECK_MV(mx,my)

  962.             }

  963.         }

  964.     }

  965. 

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

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

  968. 

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

  970.     *mx_ptr= best[0];

  971.     *my_ptr= best[1];

  972. 

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

  974.     return dmin;

  975. }

  976. 

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

  978. int ff_epzs_motion_search(MpegEncContext *s, int *mx_ptr, int *my_ptr,

  979.                           int P[10][2], int src_index, int ref_index,

  980.                           int16_t (*last_mv)[2], int ref_mv_scale,

  981.                           int size, int h)

  982. {

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

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

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

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

  987. //    case FLAG_QPEL:

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

  989.     }else{

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

  991.     }

  992. }

  993. 

  994. static int epzs_motion_search4(MpegEncContext * s,

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

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

  997.                              int ref_mv_scale)

  998. {

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

  1000.     int best[2]={0, 0};

  1001.     int d, dmin;

  1002.     unsigned map_generation;

  1003.     const int penalty_factor= c->penalty_factor;

  1004.     const int size=1;

  1005.     const int h=8;

  1006.     const int ref_mv_stride= s->mb_stride;

  1007.     const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;

  1008.     me_cmp_func cmpf, chroma_cmpf;

  1009.     LOAD_COMMON

  1010.     int flags= c->flags;

  1011.     LOAD_COMMON2

  1012. 

  1013.     cmpf= s->dsp.me_cmp[size];

  1014.     chroma_cmpf= s->dsp.me_cmp[size+1];

  1015. 

  1016.     map_generation= update_map_generation(c);

  1017. 

  1018.     dmin = 1000000;

  1019. //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);

  1020.     /* first line */

  1021.     if (s->first_slice_line) {

  1022.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1023.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1024.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1025.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1026.     }else{

  1027.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1028.         //FIXME try some early stop

  1029.         CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)

  1030.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1031.         CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)

  1032.         CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)

  1033.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

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

  1035.     }

  1036.     if(dmin>64*4){

  1037.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,

  1038.                         (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)

  1039.         if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line

  1040.             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,

  1041.                             (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)

  1042.     }

  1043. 

  1044.     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  1045. 

  1046.     *mx_ptr= best[0];

  1047.     *my_ptr= best[1];

  1048. 

  1049. //    printf("%d %d %d \n", best[0], best[1], dmin);

  1050.     return dmin;

  1051. }

  1052. 

  1053. //try to merge with above FIXME (needs PSNR test)

  1054. static int epzs_motion_search2(MpegEncContext * s,

  1055.                              int *mx_ptr, int *my_ptr, int P[10][2],

  1056.                              int src_index, int ref_index, int16_t (*last_mv)[2],

  1057.                              int ref_mv_scale)

  1058. {

  1059.     MotionEstContext * const c= &s->me;

  1060.     int best[2]={0, 0};

  1061.     int d, dmin;

  1062.     unsigned map_generation;

  1063.     const int penalty_factor= c->penalty_factor;

  1064.     const int size=0; //FIXME pass as arg

  1065.     const int h=8;

  1066.     const int ref_mv_stride= s->mb_stride;

  1067.     const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;

  1068.     me_cmp_func cmpf, chroma_cmpf;

  1069.     LOAD_COMMON

  1070.     int flags= c->flags;

  1071.     LOAD_COMMON2

  1072. 

  1073.     cmpf= s->dsp.me_cmp[size];

  1074.     chroma_cmpf= s->dsp.me_cmp[size+1];

  1075. 

  1076.     map_generation= update_map_generation(c);

  1077. 

  1078.     dmin = 1000000;

  1079. //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);

  1080.     /* first line */

  1081.     if (s->first_slice_line) {

  1082.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1083.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1084.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1085.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1086.     }else{

  1087.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1088.         //FIXME try some early stop

  1089.         CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)

  1090.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1091.         CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)

  1092.         CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)

  1093.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1094.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1095.     }

  1096.     if(dmin>64*4){

  1097.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,

  1098.                         (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)

  1099.         if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line

  1100.             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,

  1101.                             (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)

  1102.     }

  1103. 

  1104.     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  1105. 

  1106.     *mx_ptr= best[0];

  1107.     *my_ptr= best[1];

  1108. 

  1109. //    printf("%d %d %d \n", best[0], best[1], dmin);

  1110.     return dmin;

  1111. }