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.
44689 Commits
32 Branches
824MB
Tree: c5714b51aa
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'c5714b51aa'
${ noResults }
FFmpeg/libavcodec/motion_est_template.c

1109 lines
39KB
Raw Blame History 

  1. /*

  2.  * Motion estimation

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

  4.  *

  5.  * This file is part of Libav.

  6.  *

  7.  * Libav 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.  * Libav 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 Libav; 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. #include "mpegvideo.h"

  28. 

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

  30. #define LOAD_COMMON\

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

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

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

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

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

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

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

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

  39. 

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

  41. {\

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

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

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

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

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

  47. }

  48. 

  49. static int hpel_motion_search(MpegEncContext * s,

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

  51.                                   int src_index, int ref_index,

  52.                                   int size, int h)

  53. {

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

  55.     const int mx = *mx_ptr;

  56.     const int my = *my_ptr;

  57.     const int penalty_factor= c->sub_penalty_factor;

  58.     me_cmp_func cmp_sub, chroma_cmp_sub;

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

  60. 

  61.     LOAD_COMMON

  62.     int flags= c->sub_flags;

  63. 

  64.  //FIXME factorize

  65. 

  66.     cmp_sub        = s->mecc.me_sub_cmp[size];

  67.     chroma_cmp_sub = s->mecc.me_sub_cmp[size + 1];

  68. 

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

  70.         *mx_ptr = 0;

  71.         *my_ptr = 0;

  72.         return dmin;

  73.     }

  74. 

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

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

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

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

  79.     }

  80. 

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

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

  83.         int d= dmin;

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

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

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

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

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

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

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

  93. 

  94.         unsigned key;

  95.         unsigned map_generation= c->map_generation;

  96. #ifndef NDEBUG

  97.         uint32_t *map= c->map;

  98. #endif

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

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

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

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

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

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

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

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

  107.         if(t<=b){

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

  109.             if(l<=r){

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

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

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

  113.                 }else{

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

  115.                 }

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

  117.             }else{

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

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

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

  121.                 }else{

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

  123.                 }

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

  125.             }

  126.         }else{

  127.             if(l<=r){

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

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

  130.                 }else{

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

  132.                 }

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

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

  135.             }else{

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

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

  138.                 }else{

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

  140.                 }

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

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

  143.             }

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

  145.         }

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

  147.     }

  148. 

  149.     *mx_ptr = bx;

  150.     *my_ptr = by;

  151. 

  152.     return dmin;

  153. }

  154. 

  155. static int no_sub_motion_search(MpegEncContext * s,

  156.           int *mx_ptr, int *my_ptr, int dmin,

  157.                                   int src_index, int ref_index,

  158.                                   int size, int h)

  159. {

  160.     (*mx_ptr)<<=1;

  161.     (*my_ptr)<<=1;

  162.     return dmin;

  163. }

  164. 

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

  166.                                int src_index, int ref_index, int size,

  167.                                int h, int add_rate)

  168. {

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

  170.     const int penalty_factor= c->mb_penalty_factor;

  171.     const int flags= c->mb_flags;

  172.     const int qpel= flags & FLAG_QPEL;

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

  174.     me_cmp_func cmp_sub, chroma_cmp_sub;

  175.     int d;

  176. 

  177.     LOAD_COMMON

  178. 

  179.  //FIXME factorize

  180. 

  181.     cmp_sub        = s->mecc.mb_cmp[size];

  182.     chroma_cmp_sub = s->mecc.mb_cmp[size + 1];

  183. 

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

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

  186. 

  187.     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);

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

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

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

  191. 

  192.     return d;

  193. }

  194. 

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

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

  197. {

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

  199. }

  200. 

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

  202. {\

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

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

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

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

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

  208. }

  209. 

  210. static int qpel_motion_search(MpegEncContext * s,

  211.                                   int *mx_ptr, int *my_ptr, int dmin,

  212.                                   int src_index, int ref_index,

  213.                                   int size, int h)

  214. {

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

  216.     const int mx = *mx_ptr;

  217.     const int my = *my_ptr;

  218.     const int penalty_factor= c->sub_penalty_factor;

  219.     const unsigned map_generation = c->map_generation;

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

  221.     uint32_t *map= c->map;

  222.     me_cmp_func cmpf, chroma_cmpf;

  223.     me_cmp_func cmp_sub, chroma_cmp_sub;

  224. 

  225.     LOAD_COMMON

  226.     int flags= c->sub_flags;

  227. 

  228.     cmpf        = s->mecc.me_cmp[size];

  229.     chroma_cmpf = s->mecc.me_cmp[size + 1]; // FIXME: factorize

  230.  //FIXME factorize

  231. 

  232.     cmp_sub        = s->mecc.me_sub_cmp[size];

  233.     chroma_cmp_sub = s->mecc.me_sub_cmp[size + 1];

  234. 

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

  236.         *mx_ptr = 0;

  237.         *my_ptr = 0;

  238.         return dmin;

  239.     }

  240. 

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

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

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

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

  245.     }

  246. 

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

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

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

  250.         int d= dmin;

  251.         int i, nx, ny;

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

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

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

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

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

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

  258.         int best[8];

  259.         int best_pos[8][2];

  260. 

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

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

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

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

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

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

  267. 

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

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

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

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

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

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

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

  275.                     int i;

  276. 

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

  278. 

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

  280. 

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

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

  283. 

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

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

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

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

  288.                             best[i]= score;

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

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

  291.                             break;

  292.                         }

  293.                     }

  294.                 }

  295.             }

  296.         }else{

  297.             int tl;

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

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

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

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

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

  303.             int cxy;

  304. 

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

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

  307.             }else{

  308.                 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

  309.             }

  310. 

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

  312. 

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

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

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

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

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

  318. 

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

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

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

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

  323.                     int i;

  324. 

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

  326. 

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

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

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

  330. 

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

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

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

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

  335.                             best[i]= score;

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

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

  338.                             break;

  339.                         }

  340.                     }

  341.                 }

  342.             }

  343.         }

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

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

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

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

  348.         }

  349. 

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

  351. 

  352.         *mx_ptr = bx;

  353.         *my_ptr = by;

  354.     }else{

  355.         *mx_ptr =4*mx;

  356.         *my_ptr =4*my;

  357.     }

  358. 

  359.     return dmin;

  360. }

  361. 

  362. 

  363. #define CHECK_MV(x,y)\

  364. {\

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

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

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

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

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

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

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

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

  373.         map[index]= key;\

  374.         score_map[index]= d;\

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

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

  377.     }\

  378. }

  379. 

  380. #define CHECK_CLIPPED_MV(ax,ay)\

  381. {\

  382.     const int Lx= ax;\

  383.     const int Ly= ay;\

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

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

  386.     CHECK_MV(Lx2, Ly2)\

  387. }

  388. 

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

  390. {\

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

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

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

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

  395.         map[index]= key;\

  396.         score_map[index]= d;\

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

  398.         if(d<dmin){\

  399.             best[0]=x;\

  400.             best[1]=y;\

  401.             dmin=d;\

  402.             next_dir= new_dir;\

  403.         }\

  404.     }\

  405. }

  406. 

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

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

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

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

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

  412. 

  413. #define LOAD_COMMON2\

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

  415.     const int qpel= flags&FLAG_QPEL;\

  416.     const int shift= 1+qpel;\

  417. 

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

  419.                                        int src_index, int ref_index, const int penalty_factor,

  420.                                        int size, int h, int flags)

  421. {

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

  423.     me_cmp_func cmpf, chroma_cmpf;

  424.     int next_dir=-1;

  425.     LOAD_COMMON

  426.     LOAD_COMMON2

  427.     unsigned map_generation = c->map_generation;

  428. 

  429.     cmpf        = s->mecc.me_cmp[size];

  430.     chroma_cmpf = s->mecc.me_cmp[size + 1];

  431. 

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

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

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

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

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

  437.             map[index]= key;

  438.         }

  439.     }

  440. 

  441.     for(;;){

  442.         int d;

  443.         const int dir= next_dir;

  444.         const int x= best[0];

  445.         const int y= best[1];

  446.         next_dir=-1;

  447. 

  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, const int 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->mecc.me_cmp[size];

  471.     chroma_cmpf = s->mecc.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, const int 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->mecc.me_cmp[size];

  514.     chroma_cmpf = s->mecc.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, const int 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->mecc.me_cmp[size];

  551.     chroma_cmpf = s->mecc.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, const int 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->mecc.me_cmp[size];

  590.     chroma_cmpf = s->mecc.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, const int 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->mecc.me_cmp[size];

  633.     chroma_cmpf = s->mecc.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.     if(map[index]!=key){\

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

  661.         map[index]= key;\

  662.         score_map[index]= d;\

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

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

  665.             int j=0;\

  666.             \

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

  668. \

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

  670. \

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

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

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

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

  675.             \

  676.             i=-1;\

  677.             continue;\

  678.         }\

  679.     }\

  680. }

  681. 

  682. #define MAX_SAB_SIZE ME_MAP_SIZE

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

  684.                                        int src_index, int ref_index, const int penalty_factor,

  685.                                        int size, int h, int flags)

  686. {

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

  688.     me_cmp_func cmpf, chroma_cmpf;

  689.     Minima minima[MAX_SAB_SIZE];

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

  691.     int i, j;

  692.     LOAD_COMMON

  693.     LOAD_COMMON2

  694.     unsigned map_generation = c->map_generation;

  695. 

  696.     cmpf        = s->mecc.me_cmp[size];

  697.     chroma_cmpf = s->mecc.me_cmp[size + 1];

  698. 

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

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

  701.      */

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

  703.         uint32_t key= map[i];

  704. 

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

  706. 

  707.         if ((key & (-(1 << (2 * ME_MAP_MV_BITS)))) != map_generation)

  708.             continue;

  709. 

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

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

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

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

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

  715. 

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

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

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

  719.             continue;

  720. 

  721.         minima[j].checked=0;

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

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

  724. 

  725.         j++;

  726.     }

  727. 

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

  729. 

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

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

  732.         minima[j].checked=0;

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

  734.     }

  735. 

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

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

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

  739.         int d;

  740. 

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

  742. 

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

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

  745.            continue;

  746. 

  747.         SAB_CHECK_MV(x-1, y)

  748.         SAB_CHECK_MV(x+1, y)

  749.         SAB_CHECK_MV(x  , y-1)

  750.         SAB_CHECK_MV(x  , y+1)

  751. 

  752.         minima[i].checked= 1;

  753.     }

  754. 

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

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

  757.     dmin= minima[0].height;

  758. 

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

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

  761.         int d;

  762.         // ensure that the reference samples for hpel refinement are in the map

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

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

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

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

  767.     }

  768.     return dmin;

  769. }

  770. 

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

  772.                                        int src_index, int ref_index, const int penalty_factor,

  773.                                        int size, int h, int flags)

  774. {

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

  776.     me_cmp_func cmpf, chroma_cmpf;

  777.     int dia_size;

  778.     LOAD_COMMON

  779.     LOAD_COMMON2

  780.     unsigned map_generation = c->map_generation;

  781. 

  782.     cmpf        = s->mecc.me_cmp[size];

  783.     chroma_cmpf = s->mecc.me_cmp[size + 1];

  784. 

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

  786.         int dir, start, end;

  787.         const int x= best[0];

  788.         const int y= best[1];

  789. 

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

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

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

  793.             int d;

  794. 

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

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

  797.         }

  798. 

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

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

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

  802.             int d;

  803. 

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

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

  806.         }

  807. 

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

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

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

  811.             int d;

  812. 

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

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

  815.         }

  816. 

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

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

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

  820.             int d;

  821. 

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

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

  824.         }

  825. 

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

  827.             dia_size=0;

  828.     }

  829.     return dmin;

  830. }

  831. 

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

  833.                                        int src_index, int ref_index, const int penalty_factor,

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  850.     else

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

  852. }

  853. 

  854. /**

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

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

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

  858.    optimal mv.

  859.  */

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

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

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

  863. {

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

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

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

  867.                                block currently being encoded and the position of

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

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

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

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

  872.     unsigned map_generation;

  873.     int penalty_factor;

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

  875.     const int ref_mv_xy = s->mb_x + s->mb_y * ref_mv_stride; // add to last_mv before passing FIXME

  876.     me_cmp_func cmpf, chroma_cmpf;

  877. 

  878.     LOAD_COMMON

  879.     LOAD_COMMON2

  880. 

  881.     if(c->pre_pass){

  882.         penalty_factor= c->pre_penalty_factor;

  883.         cmpf           = s->mecc.me_pre_cmp[size];

  884.         chroma_cmpf    = s->mecc.me_pre_cmp[size + 1];

  885.     }else{

  886.         penalty_factor= c->penalty_factor;

  887.         cmpf           = s->mecc.me_cmp[size];

  888.         chroma_cmpf    = s->mecc.me_cmp[size + 1];

  889.     }

  890. 

  891.     map_generation= update_map_generation(c);

  892. 

  893.     assert(cmpf);

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

  895.     map[0]= map_generation;

  896.     score_map[0]= dmin;

  897. 

  898.     //FIXME precalc first term below?

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

  900.         s->mpv_flags & FF_MPV_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.     return dmin;

  974. }

  975. 

  976. //this function is dedicated to the brain damaged gcc

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

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

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

  980.                           int size, int h)

  981. {

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

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

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

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

  986. //    case FLAG_QPEL:

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

  988.     }else{

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

  990.     }

  991. }

  992. 

  993. static int epzs_motion_search4(MpegEncContext * s,

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

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

  996.                              int ref_mv_scale)

  997. {

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

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

  1000.     int d, dmin;

  1001.     unsigned map_generation;

  1002.     const int penalty_factor= c->penalty_factor;

  1003.     const int size=1;

  1004.     const int h=8;

  1005.     const int ref_mv_stride= s->mb_stride;

  1006.     const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;

  1007.     me_cmp_func cmpf, chroma_cmpf;

  1008.     LOAD_COMMON

  1009.     int flags= c->flags;

  1010.     LOAD_COMMON2

  1011. 

  1012.     cmpf        = s->mecc.me_cmp[size];

  1013.     chroma_cmpf = s->mecc.me_cmp[size + 1];

  1014. 

  1015.     map_generation= update_map_generation(c);

  1016. 

  1017.     dmin = 1000000;

  1018. 

  1019.     /* first line */

  1020.     if (s->first_slice_line) {

  1021.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1022.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

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

  1024.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1025.     }else{

  1026.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1027.         //FIXME try some early stop

  1028.         CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)

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

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

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

  1032.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

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

  1034.     }

  1035.     if(dmin>64*4){

  1036.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,

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

  1038.         if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line

  1039.             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,

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

  1041.     }

  1042. 

  1043.     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  1044. 

  1045.     *mx_ptr= best[0];

  1046.     *my_ptr= best[1];

  1047. 

  1048.     return dmin;

  1049. }

  1050. 

  1051. //try to merge with above FIXME (needs PSNR test)

  1052. static int epzs_motion_search2(MpegEncContext * s,

  1053.                              int *mx_ptr, int *my_ptr, int P[10][2],

  1054.                              int src_index, int ref_index, int16_t (*last_mv)[2],

  1055.                              int ref_mv_scale)

  1056. {

  1057.     MotionEstContext * const c= &s->me;

  1058.     int best[2]={0, 0};

  1059.     int d, dmin;

  1060.     unsigned map_generation;

  1061.     const int penalty_factor= c->penalty_factor;

  1062.     const int size=0; //FIXME pass as arg

  1063.     const int h=8;

  1064.     const int ref_mv_stride= s->mb_stride;

  1065.     const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;

  1066.     me_cmp_func cmpf, chroma_cmpf;

  1067.     LOAD_COMMON

  1068.     int flags= c->flags;

  1069.     LOAD_COMMON2

  1070. 

  1071.     cmpf        = s->mecc.me_cmp[size];

  1072.     chroma_cmpf = s->mecc.me_cmp[size + 1];

  1073. 

  1074.     map_generation= update_map_generation(c);

  1075. 

  1076.     dmin = 1000000;

  1077. 

  1078.     /* first line */

  1079.     if (s->first_slice_line) {

  1080.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

  1081.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1082.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1083.         CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)

  1084.     }else{

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

  1086.         //FIXME try some early stop

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

  1088.         CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)

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

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

  1091.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,

  1092.                         (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)

  1093.     }

  1094.     if(dmin>64*4){

  1095.         CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,

  1096.                         (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)

  1097.         if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line

  1098.             CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,

  1099.                             (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)

  1100.     }

  1101. 

  1102.     dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);

  1103. 

  1104.     *mx_ptr= best[0];

  1105.     *my_ptr= best[1];

  1106. 

  1107.     return dmin;

  1108. }