Libav
|
00001 /* 00002 * Motion estimation 00003 * Copyright (c) 2002-2004 Michael Niedermayer 00004 * 00005 * This file is part of FFmpeg. 00006 * 00007 * FFmpeg is free software; you can redistribute it and/or 00008 * modify it under the terms of the GNU Lesser General Public 00009 * License as published by the Free Software Foundation; either 00010 * version 2.1 of the License, or (at your option) any later version. 00011 * 00012 * FFmpeg is distributed in the hope that it will be useful, 00013 * but WITHOUT ANY WARRANTY; without even the implied warranty of 00014 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 00015 * Lesser General Public License for more details. 00016 * 00017 * You should have received a copy of the GNU Lesser General Public 00018 * License along with FFmpeg; if not, write to the Free Software 00019 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 00020 */ 00021 00027 //Let us hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...) 00028 #define LOAD_COMMON\ 00029 uint32_t av_unused * const score_map= c->score_map;\ 00030 const int av_unused xmin= c->xmin;\ 00031 const int av_unused ymin= c->ymin;\ 00032 const int av_unused xmax= c->xmax;\ 00033 const int av_unused ymax= c->ymax;\ 00034 uint8_t *mv_penalty= c->current_mv_penalty;\ 00035 const int pred_x= c->pred_x;\ 00036 const int pred_y= c->pred_y;\ 00037 00038 #define CHECK_HALF_MV(dx, dy, x, y)\ 00039 {\ 00040 const int hx= 2*(x)+(dx);\ 00041 const int hy= 2*(y)+(dy);\ 00042 d= cmp_hpel(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\ 00043 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\ 00044 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\ 00045 } 00046 00047 #if 0 00048 static int hpel_motion_search)(MpegEncContext * s, 00049 int *mx_ptr, int *my_ptr, int dmin, 00050 uint8_t *ref_data[3], 00051 int size) 00052 { 00053 const int xx = 16 * s->mb_x + 8*(n&1); 00054 const int yy = 16 * s->mb_y + 8*(n>>1); 00055 const int mx = *mx_ptr; 00056 const int my = *my_ptr; 00057 const int penalty_factor= c->sub_penalty_factor; 00058 00059 LOAD_COMMON 00060 00061 // INIT; 00062 //FIXME factorize 00063 me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub; 00064 00065 if(s->no_rounding /*FIXME b_type*/){ 00066 hpel_put= &s->dsp.put_no_rnd_pixels_tab[size]; 00067 chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1]; 00068 }else{ 00069 hpel_put=& s->dsp.put_pixels_tab[size]; 00070 chroma_hpel_put= &s->dsp.put_pixels_tab[size+1]; 00071 } 00072 cmpf= s->dsp.me_cmp[size]; 00073 chroma_cmpf= s->dsp.me_cmp[size+1]; 00074 cmp_sub= s->dsp.me_sub_cmp[size]; 00075 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; 00076 00077 if(c->skip){ //FIXME somehow move up (benchmark) 00078 *mx_ptr = 0; 00079 *my_ptr = 0; 00080 return dmin; 00081 } 00082 00083 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){ 00084 CMP_HPEL(dmin, 0, 0, mx, my, size); 00085 if(mx || my) 00086 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor; 00087 } 00088 00089 if (mx > xmin && mx < xmax && 00090 my > ymin && my < ymax) { 00091 int bx=2*mx, by=2*my; 00092 int d= dmin; 00093 00094 CHECK_HALF_MV(1, 1, mx-1, my-1) 00095 CHECK_HALF_MV(0, 1, mx , my-1) 00096 CHECK_HALF_MV(1, 1, mx , my-1) 00097 CHECK_HALF_MV(1, 0, mx-1, my ) 00098 CHECK_HALF_MV(1, 0, mx , my ) 00099 CHECK_HALF_MV(1, 1, mx-1, my ) 00100 CHECK_HALF_MV(0, 1, mx , my ) 00101 CHECK_HALF_MV(1, 1, mx , my ) 00102 00103 assert(bx >= xmin*2 || bx <= xmax*2 || by >= ymin*2 || by <= ymax*2); 00104 00105 *mx_ptr = bx; 00106 *my_ptr = by; 00107 }else{ 00108 *mx_ptr =2*mx; 00109 *my_ptr =2*my; 00110 } 00111 00112 return dmin; 00113 } 00114 00115 #else 00116 static int hpel_motion_search(MpegEncContext * s, 00117 int *mx_ptr, int *my_ptr, int dmin, 00118 int src_index, int ref_index, 00119 int size, int h) 00120 { 00121 MotionEstContext * const c= &s->me; 00122 const int mx = *mx_ptr; 00123 const int my = *my_ptr; 00124 const int penalty_factor= c->sub_penalty_factor; 00125 me_cmp_func cmp_sub, chroma_cmp_sub; 00126 int bx=2*mx, by=2*my; 00127 00128 LOAD_COMMON 00129 int flags= c->sub_flags; 00130 00131 //FIXME factorize 00132 00133 cmp_sub= s->dsp.me_sub_cmp[size]; 00134 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; 00135 00136 if(c->skip){ //FIXME move out of hpel? 00137 *mx_ptr = 0; 00138 *my_ptr = 0; 00139 return dmin; 00140 } 00141 00142 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){ 00143 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags); 00144 if(mx || my || size>0) 00145 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor; 00146 } 00147 00148 if (mx > xmin && mx < xmax && 00149 my > ymin && my < ymax) { 00150 int d= dmin; 00151 const int index= (my<<ME_MAP_SHIFT) + mx; 00152 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] 00153 + (mv_penalty[bx - pred_x] + mv_penalty[by-2 - pred_y])*c->penalty_factor; 00154 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)] 00155 + (mv_penalty[bx-2 - pred_x] + mv_penalty[by - pred_y])*c->penalty_factor; 00156 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)] 00157 + (mv_penalty[bx+2 - pred_x] + mv_penalty[by - pred_y])*c->penalty_factor; 00158 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] 00159 + (mv_penalty[bx - pred_x] + mv_penalty[by+2 - pred_y])*c->penalty_factor; 00160 00161 #if 1 00162 int key; 00163 int map_generation= c->map_generation; 00164 #ifndef NDEBUG 00165 uint32_t *map= c->map; 00166 #endif 00167 key= ((my-1)<<ME_MAP_MV_BITS) + (mx) + map_generation; 00168 assert(map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key); 00169 key= ((my+1)<<ME_MAP_MV_BITS) + (mx) + map_generation; 00170 assert(map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key); 00171 key= ((my)<<ME_MAP_MV_BITS) + (mx+1) + map_generation; 00172 assert(map[(index+1)&(ME_MAP_SIZE-1)] == key); 00173 key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation; 00174 assert(map[(index-1)&(ME_MAP_SIZE-1)] == key); 00175 #endif 00176 if(t<=b){ 00177 CHECK_HALF_MV(0, 1, mx ,my-1) 00178 if(l<=r){ 00179 CHECK_HALF_MV(1, 1, mx-1, my-1) 00180 if(t+r<=b+l){ 00181 CHECK_HALF_MV(1, 1, mx , my-1) 00182 }else{ 00183 CHECK_HALF_MV(1, 1, mx-1, my ) 00184 } 00185 CHECK_HALF_MV(1, 0, mx-1, my ) 00186 }else{ 00187 CHECK_HALF_MV(1, 1, mx , my-1) 00188 if(t+l<=b+r){ 00189 CHECK_HALF_MV(1, 1, mx-1, my-1) 00190 }else{ 00191 CHECK_HALF_MV(1, 1, mx , my ) 00192 } 00193 CHECK_HALF_MV(1, 0, mx , my ) 00194 } 00195 }else{ 00196 if(l<=r){ 00197 if(t+l<=b+r){ 00198 CHECK_HALF_MV(1, 1, mx-1, my-1) 00199 }else{ 00200 CHECK_HALF_MV(1, 1, mx , my ) 00201 } 00202 CHECK_HALF_MV(1, 0, mx-1, my) 00203 CHECK_HALF_MV(1, 1, mx-1, my) 00204 }else{ 00205 if(t+r<=b+l){ 00206 CHECK_HALF_MV(1, 1, mx , my-1) 00207 }else{ 00208 CHECK_HALF_MV(1, 1, mx-1, my) 00209 } 00210 CHECK_HALF_MV(1, 0, mx , my) 00211 CHECK_HALF_MV(1, 1, mx , my) 00212 } 00213 CHECK_HALF_MV(0, 1, mx , my) 00214 } 00215 assert(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2); 00216 } 00217 00218 *mx_ptr = bx; 00219 *my_ptr = by; 00220 00221 return dmin; 00222 } 00223 #endif 00224 00225 static int no_sub_motion_search(MpegEncContext * s, 00226 int *mx_ptr, int *my_ptr, int dmin, 00227 int src_index, int ref_index, 00228 int size, int h) 00229 { 00230 (*mx_ptr)<<=1; 00231 (*my_ptr)<<=1; 00232 return dmin; 00233 } 00234 00235 inline int ff_get_mb_score(MpegEncContext * s, int mx, int my, int src_index, 00236 int ref_index, int size, int h, int add_rate) 00237 { 00238 // const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp; 00239 MotionEstContext * const c= &s->me; 00240 const int penalty_factor= c->mb_penalty_factor; 00241 const int flags= c->mb_flags; 00242 const int qpel= flags & FLAG_QPEL; 00243 const int mask= 1+2*qpel; 00244 me_cmp_func cmp_sub, chroma_cmp_sub; 00245 int d; 00246 00247 LOAD_COMMON 00248 00249 //FIXME factorize 00250 00251 cmp_sub= s->dsp.mb_cmp[size]; 00252 chroma_cmp_sub= s->dsp.mb_cmp[size+1]; 00253 00254 // assert(!c->skip); 00255 // assert(c->avctx->me_sub_cmp != c->avctx->mb_cmp); 00256 00257 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); 00258 //FIXME check cbp before adding penalty for (0,0) vector 00259 if(add_rate && (mx || my || size>0)) 00260 d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor; 00261 00262 return d; 00263 } 00264 00265 #define CHECK_QUARTER_MV(dx, dy, x, y)\ 00266 {\ 00267 const int hx= 4*(x)+(dx);\ 00268 const int hy= 4*(y)+(dy);\ 00269 d= cmp_qpel(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ 00270 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\ 00271 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\ 00272 } 00273 00274 static int qpel_motion_search(MpegEncContext * s, 00275 int *mx_ptr, int *my_ptr, int dmin, 00276 int src_index, int ref_index, 00277 int size, int h) 00278 { 00279 MotionEstContext * const c= &s->me; 00280 const int mx = *mx_ptr; 00281 const int my = *my_ptr; 00282 const int penalty_factor= c->sub_penalty_factor; 00283 const int map_generation= c->map_generation; 00284 const int subpel_quality= c->avctx->me_subpel_quality; 00285 uint32_t *map= c->map; 00286 me_cmp_func cmpf, chroma_cmpf; 00287 me_cmp_func cmp_sub, chroma_cmp_sub; 00288 00289 LOAD_COMMON 00290 int flags= c->sub_flags; 00291 00292 cmpf= s->dsp.me_cmp[size]; 00293 chroma_cmpf= s->dsp.me_cmp[size+1]; //factorize FIXME 00294 //FIXME factorize 00295 00296 cmp_sub= s->dsp.me_sub_cmp[size]; 00297 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; 00298 00299 if(c->skip){ //FIXME somehow move up (benchmark) 00300 *mx_ptr = 0; 00301 *my_ptr = 0; 00302 return dmin; 00303 } 00304 00305 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){ 00306 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags); 00307 if(mx || my || size>0) 00308 dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor; 00309 } 00310 00311 if (mx > xmin && mx < xmax && 00312 my > ymin && my < ymax) { 00313 int bx=4*mx, by=4*my; 00314 int d= dmin; 00315 int i, nx, ny; 00316 const int index= (my<<ME_MAP_SHIFT) + mx; 00317 const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)]; 00318 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]; 00319 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]; 00320 const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)]; 00321 const int c= score_map[(index )&(ME_MAP_SIZE-1)]; 00322 int best[8]; 00323 int best_pos[8][2]; 00324 00325 memset(best, 64, sizeof(int)*8); 00326 #if 1 00327 if(s->me.dia_size>=2){ 00328 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; 00329 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; 00330 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; 00331 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; 00332 00333 for(ny= -3; ny <= 3; ny++){ 00334 for(nx= -3; nx <= 3; nx++){ 00335 //FIXME this could overflow (unlikely though) 00336 const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t; 00337 const int64_t c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c; 00338 const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b; 00339 int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10; 00340 int i; 00341 00342 if((nx&3)==0 && (ny&3)==0) continue; 00343 00344 score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor; 00345 00346 // if(nx&1) score-=1024*c->penalty_factor; 00347 // if(ny&1) score-=1024*c->penalty_factor; 00348 00349 for(i=0; i<8; i++){ 00350 if(score < best[i]){ 00351 memmove(&best[i+1], &best[i], sizeof(int)*(7-i)); 00352 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i)); 00353 best[i]= score; 00354 best_pos[i][0]= nx + 4*mx; 00355 best_pos[i][1]= ny + 4*my; 00356 break; 00357 } 00358 } 00359 } 00360 } 00361 }else{ 00362 int tl; 00363 //FIXME this could overflow (unlikely though) 00364 const int cx = 4*(r - l); 00365 const int cx2= r + l - 2*c; 00366 const int cy = 4*(b - t); 00367 const int cy2= b + t - 2*c; 00368 int cxy; 00369 00370 if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME 00371 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; 00372 }else{ 00373 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 00374 } 00375 00376 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c; 00377 00378 assert(16*cx2 + 4*cx + 32*c == 32*r); 00379 assert(16*cx2 - 4*cx + 32*c == 32*l); 00380 assert(16*cy2 + 4*cy + 32*c == 32*b); 00381 assert(16*cy2 - 4*cy + 32*c == 32*t); 00382 assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl); 00383 00384 for(ny= -3; ny <= 3; ny++){ 00385 for(nx= -3; nx <= 3; nx++){ 00386 //FIXME this could overflow (unlikely though) 00387 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor 00388 int i; 00389 00390 if((nx&3)==0 && (ny&3)==0) continue; 00391 00392 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor; 00393 // if(nx&1) score-=32*c->penalty_factor; 00394 // if(ny&1) score-=32*c->penalty_factor; 00395 00396 for(i=0; i<8; i++){ 00397 if(score < best[i]){ 00398 memmove(&best[i+1], &best[i], sizeof(int)*(7-i)); 00399 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i)); 00400 best[i]= score; 00401 best_pos[i][0]= nx + 4*mx; 00402 best_pos[i][1]= ny + 4*my; 00403 break; 00404 } 00405 } 00406 } 00407 } 00408 } 00409 for(i=0; i<subpel_quality; i++){ 00410 nx= best_pos[i][0]; 00411 ny= best_pos[i][1]; 00412 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2) 00413 } 00414 00415 #if 0 00416 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; 00417 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; 00418 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; 00419 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; 00420 // if(l < r && l < t && l < b && l < tl && l < bl && l < tr && l < br && bl < tl){ 00421 if(tl<br){ 00422 00423 // nx= FFMAX(4*mx - bx, bx - 4*mx); 00424 // ny= FFMAX(4*my - by, by - 4*my); 00425 00426 static int stats[7][7], count; 00427 count++; 00428 stats[4*mx - bx + 3][4*my - by + 3]++; 00429 if(256*256*256*64 % count ==0){ 00430 for(i=0; i<49; i++){ 00431 if((i%7)==0) printf("\n"); 00432 printf("%6d ", stats[0][i]); 00433 } 00434 printf("\n"); 00435 } 00436 } 00437 #endif 00438 #else 00439 00440 CHECK_QUARTER_MV(2, 2, mx-1, my-1) 00441 CHECK_QUARTER_MV(0, 2, mx , my-1) 00442 CHECK_QUARTER_MV(2, 2, mx , my-1) 00443 CHECK_QUARTER_MV(2, 0, mx , my ) 00444 CHECK_QUARTER_MV(2, 2, mx , my ) 00445 CHECK_QUARTER_MV(0, 2, mx , my ) 00446 CHECK_QUARTER_MV(2, 2, mx-1, my ) 00447 CHECK_QUARTER_MV(2, 0, mx-1, my ) 00448 00449 nx= bx; 00450 ny= by; 00451 00452 for(i=0; i<8; i++){ 00453 int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1}; 00454 int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1}; 00455 CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>2, (ny + oy[i])>>2) 00456 } 00457 #endif 00458 #if 0 00459 //outer ring 00460 CHECK_QUARTER_MV(1, 3, mx-1, my-1) 00461 CHECK_QUARTER_MV(1, 2, mx-1, my-1) 00462 CHECK_QUARTER_MV(1, 1, mx-1, my-1) 00463 CHECK_QUARTER_MV(2, 1, mx-1, my-1) 00464 CHECK_QUARTER_MV(3, 1, mx-1, my-1) 00465 CHECK_QUARTER_MV(0, 1, mx , my-1) 00466 CHECK_QUARTER_MV(1, 1, mx , my-1) 00467 CHECK_QUARTER_MV(2, 1, mx , my-1) 00468 CHECK_QUARTER_MV(3, 1, mx , my-1) 00469 CHECK_QUARTER_MV(3, 2, mx , my-1) 00470 CHECK_QUARTER_MV(3, 3, mx , my-1) 00471 CHECK_QUARTER_MV(3, 0, mx , my ) 00472 CHECK_QUARTER_MV(3, 1, mx , my ) 00473 CHECK_QUARTER_MV(3, 2, mx , my ) 00474 CHECK_QUARTER_MV(3, 3, mx , my ) 00475 CHECK_QUARTER_MV(2, 3, mx , my ) 00476 CHECK_QUARTER_MV(1, 3, mx , my ) 00477 CHECK_QUARTER_MV(0, 3, mx , my ) 00478 CHECK_QUARTER_MV(3, 3, mx-1, my ) 00479 CHECK_QUARTER_MV(2, 3, mx-1, my ) 00480 CHECK_QUARTER_MV(1, 3, mx-1, my ) 00481 CHECK_QUARTER_MV(1, 2, mx-1, my ) 00482 CHECK_QUARTER_MV(1, 1, mx-1, my ) 00483 CHECK_QUARTER_MV(1, 0, mx-1, my ) 00484 #endif 00485 assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4); 00486 00487 *mx_ptr = bx; 00488 *my_ptr = by; 00489 }else{ 00490 *mx_ptr =4*mx; 00491 *my_ptr =4*my; 00492 } 00493 00494 return dmin; 00495 } 00496 00497 00498 #define CHECK_MV(x,y)\ 00499 {\ 00500 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\ 00501 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\ 00502 assert((x) >= xmin);\ 00503 assert((x) <= xmax);\ 00504 assert((y) >= ymin);\ 00505 assert((y) <= ymax);\ 00506 /*printf("check_mv %d %d\n", x, y);*/\ 00507 if(map[index]!=key){\ 00508 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ 00509 map[index]= key;\ 00510 score_map[index]= d;\ 00511 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\ 00512 /*printf("score:%d\n", d);*/\ 00513 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\ 00514 }\ 00515 } 00516 00517 #define CHECK_CLIPPED_MV(ax,ay)\ 00518 {\ 00519 const int Lx= ax;\ 00520 const int Ly= ay;\ 00521 const int Lx2= FFMAX(xmin, FFMIN(Lx, xmax));\ 00522 const int Ly2= FFMAX(ymin, FFMIN(Ly, ymax));\ 00523 CHECK_MV(Lx2, Ly2)\ 00524 } 00525 00526 #define CHECK_MV_DIR(x,y,new_dir)\ 00527 {\ 00528 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\ 00529 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\ 00530 /*printf("check_mv_dir %d %d %d\n", x, y, new_dir);*/\ 00531 if(map[index]!=key){\ 00532 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ 00533 map[index]= key;\ 00534 score_map[index]= d;\ 00535 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\ 00536 /*printf("score:%d\n", d);*/\ 00537 if(d<dmin){\ 00538 best[0]=x;\ 00539 best[1]=y;\ 00540 dmin=d;\ 00541 next_dir= new_dir;\ 00542 }\ 00543 }\ 00544 } 00545 00546 #define check(x,y,S,v)\ 00547 if( (x)<(xmin<<(S)) ) printf("%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\ 00548 if( (x)>(xmax<<(S)) ) printf("%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\ 00549 if( (y)<(ymin<<(S)) ) printf("%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\ 00550 if( (y)>(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\ 00551 00552 #define LOAD_COMMON2\ 00553 uint32_t *map= c->map;\ 00554 const int qpel= flags&FLAG_QPEL;\ 00555 const int shift= 1+qpel;\ 00556 00557 static av_always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin, 00558 int src_index, int ref_index, int const penalty_factor, 00559 int size, int h, int flags) 00560 { 00561 MotionEstContext * const c= &s->me; 00562 me_cmp_func cmpf, chroma_cmpf; 00563 int next_dir=-1; 00564 LOAD_COMMON 00565 LOAD_COMMON2 00566 int map_generation= c->map_generation; 00567 00568 cmpf= s->dsp.me_cmp[size]; 00569 chroma_cmpf= s->dsp.me_cmp[size+1]; 00570 00571 { /* ensure that the best point is in the MAP as h/qpel refinement needs it */ 00572 const int key= (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation; 00573 const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1); 00574 if(map[index]!=key){ //this will be executed only very rarey 00575 score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags); 00576 map[index]= key; 00577 } 00578 } 00579 00580 for(;;){ 00581 int d; 00582 const int dir= next_dir; 00583 const int x= best[0]; 00584 const int y= best[1]; 00585 next_dir=-1; 00586 00587 //printf("%d", dir); 00588 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0) 00589 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1) 00590 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2) 00591 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3) 00592 00593 if(next_dir==-1){ 00594 return dmin; 00595 } 00596 } 00597 } 00598 00599 static int funny_diamond_search(MpegEncContext * s, int *best, int dmin, 00600 int src_index, int ref_index, int const penalty_factor, 00601 int size, int h, int flags) 00602 { 00603 MotionEstContext * const c= &s->me; 00604 me_cmp_func cmpf, chroma_cmpf; 00605 int dia_size; 00606 LOAD_COMMON 00607 LOAD_COMMON2 00608 int map_generation= c->map_generation; 00609 00610 cmpf= s->dsp.me_cmp[size]; 00611 chroma_cmpf= s->dsp.me_cmp[size+1]; 00612 00613 for(dia_size=1; dia_size<=4; dia_size++){ 00614 int dir; 00615 const int x= best[0]; 00616 const int y= best[1]; 00617 00618 if(dia_size&(dia_size-1)) continue; 00619 00620 if( x + dia_size > xmax 00621 || x - dia_size < xmin 00622 || y + dia_size > ymax 00623 || y - dia_size < ymin) 00624 continue; 00625 00626 for(dir= 0; dir<dia_size; dir+=2){ 00627 int d; 00628 00629 CHECK_MV(x + dir , y + dia_size - dir); 00630 CHECK_MV(x + dia_size - dir, y - dir ); 00631 CHECK_MV(x - dir , y - dia_size + dir); 00632 CHECK_MV(x - dia_size + dir, y + dir ); 00633 } 00634 00635 if(x!=best[0] || y!=best[1]) 00636 dia_size=0; 00637 #if 0 00638 { 00639 int dx, dy, i; 00640 static int stats[8*8]; 00641 dx= FFABS(x-best[0]); 00642 dy= FFABS(y-best[1]); 00643 if(dy>dx){ 00644 dx^=dy; dy^=dx; dx^=dy; 00645 } 00646 stats[dy*8 + dx] ++; 00647 if(256*256*256*64 % (stats[0]+1)==0){ 00648 for(i=0; i<64; i++){ 00649 if((i&7)==0) printf("\n"); 00650 printf("%8d ", stats[i]); 00651 } 00652 printf("\n"); 00653 } 00654 } 00655 #endif 00656 } 00657 return dmin; 00658 } 00659 00660 static int hex_search(MpegEncContext * s, int *best, int dmin, 00661 int src_index, int ref_index, int const penalty_factor, 00662 int size, int h, int flags, int dia_size) 00663 { 00664 MotionEstContext * const c= &s->me; 00665 me_cmp_func cmpf, chroma_cmpf; 00666 LOAD_COMMON 00667 LOAD_COMMON2 00668 int map_generation= c->map_generation; 00669 int x,y,d; 00670 const int dec= dia_size & (dia_size-1); 00671 00672 cmpf= s->dsp.me_cmp[size]; 00673 chroma_cmpf= s->dsp.me_cmp[size+1]; 00674 00675 for(;dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){ 00676 do{ 00677 x= best[0]; 00678 y= best[1]; 00679 00680 CHECK_CLIPPED_MV(x -dia_size , y); 00681 CHECK_CLIPPED_MV(x+ dia_size , y); 00682 CHECK_CLIPPED_MV(x+( dia_size>>1), y+dia_size); 00683 CHECK_CLIPPED_MV(x+( dia_size>>1), y-dia_size); 00684 if(dia_size>1){ 00685 CHECK_CLIPPED_MV(x+(-dia_size>>1), y+dia_size); 00686 CHECK_CLIPPED_MV(x+(-dia_size>>1), y-dia_size); 00687 } 00688 }while(best[0] != x || best[1] != y); 00689 } 00690 00691 return dmin; 00692 } 00693 00694 static int l2s_dia_search(MpegEncContext * s, int *best, int dmin, 00695 int src_index, int ref_index, int const penalty_factor, 00696 int size, int h, int flags) 00697 { 00698 MotionEstContext * const c= &s->me; 00699 me_cmp_func cmpf, chroma_cmpf; 00700 LOAD_COMMON 00701 LOAD_COMMON2 00702 int map_generation= c->map_generation; 00703 int x,y,i,d; 00704 int dia_size= c->dia_size&0xFF; 00705 const int dec= dia_size & (dia_size-1); 00706 static const int hex[8][2]={{-2, 0}, {-1,-1}, { 0,-2}, { 1,-1}, 00707 { 2, 0}, { 1, 1}, { 0, 2}, {-1, 1}}; 00708 00709 cmpf= s->dsp.me_cmp[size]; 00710 chroma_cmpf= s->dsp.me_cmp[size+1]; 00711 00712 for(; dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){ 00713 do{ 00714 x= best[0]; 00715 y= best[1]; 00716 for(i=0; i<8; i++){ 00717 CHECK_CLIPPED_MV(x+hex[i][0]*dia_size, y+hex[i][1]*dia_size); 00718 } 00719 }while(best[0] != x || best[1] != y); 00720 } 00721 00722 x= best[0]; 00723 y= best[1]; 00724 CHECK_CLIPPED_MV(x+1, y); 00725 CHECK_CLIPPED_MV(x, y+1); 00726 CHECK_CLIPPED_MV(x-1, y); 00727 CHECK_CLIPPED_MV(x, y-1); 00728 00729 return dmin; 00730 } 00731 00732 static int umh_search(MpegEncContext * s, int *best, int dmin, 00733 int src_index, int ref_index, int const penalty_factor, 00734 int size, int h, int flags) 00735 { 00736 MotionEstContext * const c= &s->me; 00737 me_cmp_func cmpf, chroma_cmpf; 00738 LOAD_COMMON 00739 LOAD_COMMON2 00740 int map_generation= c->map_generation; 00741 int x,y,x2,y2, i, j, d; 00742 const int dia_size= c->dia_size&0xFE; 00743 static const int hex[16][2]={{-4,-2}, {-4,-1}, {-4, 0}, {-4, 1}, {-4, 2}, 00744 { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2}, 00745 {-2, 3}, { 0, 4}, { 2, 3}, 00746 {-2,-3}, { 0,-4}, { 2,-3},}; 00747 00748 cmpf= s->dsp.me_cmp[size]; 00749 chroma_cmpf= s->dsp.me_cmp[size+1]; 00750 00751 x= best[0]; 00752 y= best[1]; 00753 for(x2=FFMAX(x-dia_size+1, xmin); x2<=FFMIN(x+dia_size-1,xmax); x2+=2){ 00754 CHECK_MV(x2, y); 00755 } 00756 for(y2=FFMAX(y-dia_size/2+1, ymin); y2<=FFMIN(y+dia_size/2-1,ymax); y2+=2){ 00757 CHECK_MV(x, y2); 00758 } 00759 00760 x= best[0]; 00761 y= best[1]; 00762 for(y2=FFMAX(y-2, ymin); y2<=FFMIN(y+2,ymax); y2++){ 00763 for(x2=FFMAX(x-2, xmin); x2<=FFMIN(x+2,xmax); x2++){ 00764 CHECK_MV(x2, y2); 00765 } 00766 } 00767 00768 //FIXME prevent the CLIP stuff 00769 00770 for(j=1; j<=dia_size/4; j++){ 00771 for(i=0; i<16; i++){ 00772 CHECK_CLIPPED_MV(x+hex[i][0]*j, y+hex[i][1]*j); 00773 } 00774 } 00775 00776 return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, 2); 00777 } 00778 00779 static int full_search(MpegEncContext * s, int *best, int dmin, 00780 int src_index, int ref_index, int const penalty_factor, 00781 int size, int h, int flags) 00782 { 00783 MotionEstContext * const c= &s->me; 00784 me_cmp_func cmpf, chroma_cmpf; 00785 LOAD_COMMON 00786 LOAD_COMMON2 00787 int map_generation= c->map_generation; 00788 int x,y, d; 00789 const int dia_size= c->dia_size&0xFF; 00790 00791 cmpf= s->dsp.me_cmp[size]; 00792 chroma_cmpf= s->dsp.me_cmp[size+1]; 00793 00794 for(y=FFMAX(-dia_size, ymin); y<=FFMIN(dia_size,ymax); y++){ 00795 for(x=FFMAX(-dia_size, xmin); x<=FFMIN(dia_size,xmax); x++){ 00796 CHECK_MV(x, y); 00797 } 00798 } 00799 00800 x= best[0]; 00801 y= best[1]; 00802 d= dmin; 00803 CHECK_CLIPPED_MV(x , y); 00804 CHECK_CLIPPED_MV(x+1, y); 00805 CHECK_CLIPPED_MV(x, y+1); 00806 CHECK_CLIPPED_MV(x-1, y); 00807 CHECK_CLIPPED_MV(x, y-1); 00808 best[0]= x; 00809 best[1]= y; 00810 00811 return d; 00812 } 00813 00814 #define SAB_CHECK_MV(ax,ay)\ 00815 {\ 00816 const int key= ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\ 00817 const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\ 00818 /*printf("sab check %d %d\n", ax, ay);*/\ 00819 if(map[index]!=key){\ 00820 d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ 00821 map[index]= key;\ 00822 score_map[index]= d;\ 00823 d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\ 00824 /*printf("score: %d\n", d);*/\ 00825 if(d < minima[minima_count-1].height){\ 00826 int j=0;\ 00827 \ 00828 while(d >= minima[j].height) j++;\ 00829 \ 00830 memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\ 00831 \ 00832 minima[j].checked= 0;\ 00833 minima[j].height= d;\ 00834 minima[j].x= ax;\ 00835 minima[j].y= ay;\ 00836 \ 00837 i=-1;\ 00838 continue;\ 00839 }\ 00840 }\ 00841 } 00842 00843 #define MAX_SAB_SIZE ME_MAP_SIZE 00844 static int sab_diamond_search(MpegEncContext * s, int *best, int dmin, 00845 int src_index, int ref_index, int const penalty_factor, 00846 int size, int h, int flags) 00847 { 00848 MotionEstContext * const c= &s->me; 00849 me_cmp_func cmpf, chroma_cmpf; 00850 Minima minima[MAX_SAB_SIZE]; 00851 const int minima_count= FFABS(c->dia_size); 00852 int i, j; 00853 LOAD_COMMON 00854 LOAD_COMMON2 00855 int map_generation= c->map_generation; 00856 00857 cmpf= s->dsp.me_cmp[size]; 00858 chroma_cmpf= s->dsp.me_cmp[size+1]; 00859 00860 /*Note j<MAX_SAB_SIZE is needed if MAX_SAB_SIZE < ME_MAP_SIZE as j can 00861 become larger due to MVs overflowing their ME_MAP_MV_BITS bits space in map 00862 */ 00863 for(j=i=0; i<ME_MAP_SIZE && j<MAX_SAB_SIZE; i++){ 00864 uint32_t key= map[i]; 00865 00866 key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1)); 00867 00868 if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue; 00869 00870 minima[j].height= score_map[i]; 00871 minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS; 00872 minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1); 00873 minima[j].x-= (1<<(ME_MAP_MV_BITS-1)); 00874 minima[j].y-= (1<<(ME_MAP_MV_BITS-1)); 00875 00876 // all entries in map should be in range except if the mv overflows their ME_MAP_MV_BITS bits space 00877 if( minima[j].x > xmax || minima[j].x < xmin 00878 || minima[j].y > ymax || minima[j].y < ymin) 00879 continue; 00880 00881 minima[j].checked=0; 00882 if(minima[j].x || minima[j].y) 00883 minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor; 00884 00885 j++; 00886 } 00887 00888 qsort(minima, j, sizeof(Minima), minima_cmp); 00889 00890 for(; j<minima_count; j++){ 00891 minima[j].height=256*256*256*64; 00892 minima[j].checked=0; 00893 minima[j].x= minima[j].y=0; 00894 } 00895 00896 for(i=0; i<minima_count; i++){ 00897 const int x= minima[i].x; 00898 const int y= minima[i].y; 00899 int d; 00900 00901 if(minima[i].checked) continue; 00902 00903 if( x >= xmax || x <= xmin 00904 || y >= ymax || y <= ymin) 00905 continue; 00906 00907 SAB_CHECK_MV(x-1, y) 00908 SAB_CHECK_MV(x+1, y) 00909 SAB_CHECK_MV(x , y-1) 00910 SAB_CHECK_MV(x , y+1) 00911 00912 minima[i].checked= 1; 00913 } 00914 00915 best[0]= minima[0].x; 00916 best[1]= minima[0].y; 00917 dmin= minima[0].height; 00918 00919 if( best[0] < xmax && best[0] > xmin 00920 && best[1] < ymax && best[1] > ymin){ 00921 int d; 00922 //ensure that the refernece samples for hpel refinement are in the map 00923 CHECK_MV(best[0]-1, best[1]) 00924 CHECK_MV(best[0]+1, best[1]) 00925 CHECK_MV(best[0], best[1]-1) 00926 CHECK_MV(best[0], best[1]+1) 00927 } 00928 return dmin; 00929 } 00930 00931 static int var_diamond_search(MpegEncContext * s, int *best, int dmin, 00932 int src_index, int ref_index, int const penalty_factor, 00933 int size, int h, int flags) 00934 { 00935 MotionEstContext * const c= &s->me; 00936 me_cmp_func cmpf, chroma_cmpf; 00937 int dia_size; 00938 LOAD_COMMON 00939 LOAD_COMMON2 00940 int map_generation= c->map_generation; 00941 00942 cmpf= s->dsp.me_cmp[size]; 00943 chroma_cmpf= s->dsp.me_cmp[size+1]; 00944 00945 for(dia_size=1; dia_size<=c->dia_size; dia_size++){ 00946 int dir, start, end; 00947 const int x= best[0]; 00948 const int y= best[1]; 00949 00950 start= FFMAX(0, y + dia_size - ymax); 00951 end = FFMIN(dia_size, xmax - x + 1); 00952 for(dir= start; dir<end; dir++){ 00953 int d; 00954 00955 //check(x + dir,y + dia_size - dir,0, a0) 00956 CHECK_MV(x + dir , y + dia_size - dir); 00957 } 00958 00959 start= FFMAX(0, x + dia_size - xmax); 00960 end = FFMIN(dia_size, y - ymin + 1); 00961 for(dir= start; dir<end; dir++){ 00962 int d; 00963 00964 //check(x + dia_size - dir, y - dir,0, a1) 00965 CHECK_MV(x + dia_size - dir, y - dir ); 00966 } 00967 00968 start= FFMAX(0, -y + dia_size + ymin ); 00969 end = FFMIN(dia_size, x - xmin + 1); 00970 for(dir= start; dir<end; dir++){ 00971 int d; 00972 00973 //check(x - dir,y - dia_size + dir,0, a2) 00974 CHECK_MV(x - dir , y - dia_size + dir); 00975 } 00976 00977 start= FFMAX(0, -x + dia_size + xmin ); 00978 end = FFMIN(dia_size, ymax - y + 1); 00979 for(dir= start; dir<end; dir++){ 00980 int d; 00981 00982 //check(x - dia_size + dir, y + dir,0, a3) 00983 CHECK_MV(x - dia_size + dir, y + dir ); 00984 } 00985 00986 if(x!=best[0] || y!=best[1]) 00987 dia_size=0; 00988 #if 0 00989 { 00990 int dx, dy, i; 00991 static int stats[8*8]; 00992 dx= FFABS(x-best[0]); 00993 dy= FFABS(y-best[1]); 00994 stats[dy*8 + dx] ++; 00995 if(256*256*256*64 % (stats[0]+1)==0){ 00996 for(i=0; i<64; i++){ 00997 if((i&7)==0) printf("\n"); 00998 printf("%6d ", stats[i]); 00999 } 01000 printf("\n"); 01001 } 01002 } 01003 #endif 01004 } 01005 return dmin; 01006 } 01007 01008 static av_always_inline int diamond_search(MpegEncContext * s, int *best, int dmin, 01009 int src_index, int ref_index, int const penalty_factor, 01010 int size, int h, int flags){ 01011 MotionEstContext * const c= &s->me; 01012 if(c->dia_size==-1) 01013 return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01014 else if(c->dia_size<-1) 01015 return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01016 else if(c->dia_size<2) 01017 return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01018 else if(c->dia_size>1024) 01019 return full_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01020 else if(c->dia_size>768) 01021 return umh_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01022 else if(c->dia_size>512) 01023 return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, c->dia_size&0xFF); 01024 else if(c->dia_size>256) 01025 return l2s_dia_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01026 else 01027 return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01028 } 01029 01036 static av_always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr, 01037 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2], 01038 int ref_mv_scale, int flags, int size, int h) 01039 { 01040 MotionEstContext * const c= &s->me; 01041 int best[2]={0, 0}; 01045 int d; 01046 int dmin; 01048 int map_generation; 01049 int penalty_factor; 01050 const int ref_mv_stride= s->mb_stride; //pass as arg FIXME 01051 const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME 01052 me_cmp_func cmpf, chroma_cmpf; 01053 01054 LOAD_COMMON 01055 LOAD_COMMON2 01056 01057 if(c->pre_pass){ 01058 penalty_factor= c->pre_penalty_factor; 01059 cmpf= s->dsp.me_pre_cmp[size]; 01060 chroma_cmpf= s->dsp.me_pre_cmp[size+1]; 01061 }else{ 01062 penalty_factor= c->penalty_factor; 01063 cmpf= s->dsp.me_cmp[size]; 01064 chroma_cmpf= s->dsp.me_cmp[size+1]; 01065 } 01066 01067 map_generation= update_map_generation(c); 01068 01069 assert(cmpf); 01070 dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags); 01071 map[0]= map_generation; 01072 score_map[0]= dmin; 01073 01074 //FIXME precalc first term below? 01075 if((s->pict_type == FF_B_TYPE && !(c->flags & FLAG_DIRECT)) || s->flags&CODEC_FLAG_MV0) 01076 dmin += (mv_penalty[pred_x] + mv_penalty[pred_y])*penalty_factor; 01077 01078 /* first line */ 01079 if (s->first_slice_line) { 01080 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) 01081 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 01082 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) 01083 }else{ 01084 if(dmin<((h*h*s->avctx->mv0_threshold)>>8) 01085 && ( P_LEFT[0] |P_LEFT[1] 01086 |P_TOP[0] |P_TOP[1] 01087 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){ 01088 *mx_ptr= 0; 01089 *my_ptr= 0; 01090 c->skip=1; 01091 return dmin; 01092 } 01093 CHECK_MV( P_MEDIAN[0] >>shift , P_MEDIAN[1] >>shift) 01094 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)-1) 01095 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)+1) 01096 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)-1, (P_MEDIAN[1]>>shift) ) 01097 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)+1, (P_MEDIAN[1]>>shift) ) 01098 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 01099 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) 01100 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift) 01101 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift) 01102 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) 01103 } 01104 if(dmin>h*h*4){ 01105 if(c->pre_pass){ 01106 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16, 01107 (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16) 01108 if(!s->first_slice_line) 01109 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, 01110 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) 01111 }else{ 01112 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, 01113 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16) 01114 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line 01115 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, 01116 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) 01117 } 01118 } 01119 01120 if(c->avctx->last_predictor_count){ 01121 const int count= c->avctx->last_predictor_count; 01122 const int xstart= FFMAX(0, s->mb_x - count); 01123 const int ystart= FFMAX(0, s->mb_y - count); 01124 const int xend= FFMIN(s->mb_width , s->mb_x + count + 1); 01125 const int yend= FFMIN(s->mb_height, s->mb_y + count + 1); 01126 int mb_y; 01127 01128 for(mb_y=ystart; mb_y<yend; mb_y++){ 01129 int mb_x; 01130 for(mb_x=xstart; mb_x<xend; mb_x++){ 01131 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride; 01132 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16; 01133 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16; 01134 01135 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue; 01136 CHECK_MV(mx,my) 01137 } 01138 } 01139 } 01140 01141 //check(best[0],best[1],0, b0) 01142 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01143 01144 //check(best[0],best[1],0, b1) 01145 *mx_ptr= best[0]; 01146 *my_ptr= best[1]; 01147 01148 // printf("%d %d %d \n", best[0], best[1], dmin); 01149 return dmin; 01150 } 01151 01152 //this function is dedicated to the braindamaged gcc 01153 inline int ff_epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr, 01154 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2], 01155 int ref_mv_scale, int size, int h) 01156 { 01157 MotionEstContext * const c= &s->me; 01158 //FIXME convert other functions in the same way if faster 01159 if(c->flags==0 && h==16 && size==0){ 01160 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16); 01161 // case FLAG_QPEL: 01162 // return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL); 01163 }else{ 01164 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h); 01165 } 01166 } 01167 01168 static int epzs_motion_search4(MpegEncContext * s, 01169 int *mx_ptr, int *my_ptr, int P[10][2], 01170 int src_index, int ref_index, int16_t (*last_mv)[2], 01171 int ref_mv_scale) 01172 { 01173 MotionEstContext * const c= &s->me; 01174 int best[2]={0, 0}; 01175 int d, dmin; 01176 int map_generation; 01177 const int penalty_factor= c->penalty_factor; 01178 const int size=1; 01179 const int h=8; 01180 const int ref_mv_stride= s->mb_stride; 01181 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride; 01182 me_cmp_func cmpf, chroma_cmpf; 01183 LOAD_COMMON 01184 int flags= c->flags; 01185 LOAD_COMMON2 01186 01187 cmpf= s->dsp.me_cmp[size]; 01188 chroma_cmpf= s->dsp.me_cmp[size+1]; 01189 01190 map_generation= update_map_generation(c); 01191 01192 dmin = 1000000; 01193 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax); 01194 /* first line */ 01195 if (s->first_slice_line) { 01196 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) 01197 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 01198 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) 01199 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) 01200 }else{ 01201 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) 01202 //FIXME try some early stop 01203 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift) 01204 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) 01205 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift) 01206 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) 01207 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 01208 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) 01209 } 01210 if(dmin>64*4){ 01211 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, 01212 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16) 01213 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line 01214 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, 01215 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) 01216 } 01217 01218 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01219 01220 *mx_ptr= best[0]; 01221 *my_ptr= best[1]; 01222 01223 // printf("%d %d %d \n", best[0], best[1], dmin); 01224 return dmin; 01225 } 01226 01227 //try to merge with above FIXME (needs PSNR test) 01228 static int epzs_motion_search2(MpegEncContext * s, 01229 int *mx_ptr, int *my_ptr, int P[10][2], 01230 int src_index, int ref_index, int16_t (*last_mv)[2], 01231 int ref_mv_scale) 01232 { 01233 MotionEstContext * const c= &s->me; 01234 int best[2]={0, 0}; 01235 int d, dmin; 01236 int map_generation; 01237 const int penalty_factor= c->penalty_factor; 01238 const int size=0; //FIXME pass as arg 01239 const int h=8; 01240 const int ref_mv_stride= s->mb_stride; 01241 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride; 01242 me_cmp_func cmpf, chroma_cmpf; 01243 LOAD_COMMON 01244 int flags= c->flags; 01245 LOAD_COMMON2 01246 01247 cmpf= s->dsp.me_cmp[size]; 01248 chroma_cmpf= s->dsp.me_cmp[size+1]; 01249 01250 map_generation= update_map_generation(c); 01251 01252 dmin = 1000000; 01253 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax); 01254 /* first line */ 01255 if (s->first_slice_line) { 01256 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) 01257 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 01258 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) 01259 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) 01260 }else{ 01261 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) 01262 //FIXME try some early stop 01263 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift) 01264 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) 01265 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift) 01266 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) 01267 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 01268 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) 01269 } 01270 if(dmin>64*4){ 01271 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, 01272 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16) 01273 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line 01274 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, 01275 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) 01276 } 01277 01278 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); 01279 01280 *mx_ptr= best[0]; 01281 *my_ptr= best[1]; 01282 01283 // printf("%d %d %d \n", best[0], best[1], dmin); 01284 return dmin; 01285 }