| 1 | /* |
| 2 | * Motion estimation |
| 3 | * Copyright (c) 2002-2004 Michael Niedermayer |
| 4 | * |
| 5 | * This file is part of FFmpeg. |
| 6 | * |
| 7 | * FFmpeg is free software; you can redistribute it and/or |
| 8 | * modify it under the terms of the GNU Lesser General Public |
| 9 | * License as published by the Free Software Foundation; either |
| 10 | * version 2.1 of the License, or (at your option) any later version. |
| 11 | * |
| 12 | * FFmpeg is distributed in the hope that it will be useful, |
| 13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 15 | * Lesser General Public License for more details. |
| 16 | * |
| 17 | * You should have received a copy of the GNU Lesser General Public |
| 18 | * License along with FFmpeg; if not, write to the Free Software |
| 19 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 20 | */ |
| 21 | |
| 22 | /** |
| 23 | * @file |
| 24 | * Motion estimation template. |
| 25 | */ |
| 26 | |
| 27 | #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 | #if defined(ASSERT_LEVEL) && ASSERT_LEVEL > 1 |
| 95 | unsigned key; |
| 96 | unsigned map_generation= c->map_generation; |
| 97 | key= ((my-1)<<ME_MAP_MV_BITS) + (mx) + map_generation; |
| 98 | av_assert2(c->map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key); |
| 99 | key= ((my+1)<<ME_MAP_MV_BITS) + (mx) + map_generation; |
| 100 | av_assert2(c->map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key); |
| 101 | key= ((my)<<ME_MAP_MV_BITS) + (mx+1) + map_generation; |
| 102 | av_assert2(c->map[(index+1)&(ME_MAP_SIZE-1)] == key); |
| 103 | key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation; |
| 104 | av_assert2(c->map[(index-1)&(ME_MAP_SIZE-1)] == key); |
| 105 | #endif |
| 106 | if(t<=b){ |
| 107 | CHECK_HALF_MV(0, 1, mx ,my-1) |
| 108 | if(l<=r){ |
| 109 | CHECK_HALF_MV(1, 1, mx-1, my-1) |
| 110 | if(t+r<=b+l){ |
| 111 | CHECK_HALF_MV(1, 1, mx , my-1) |
| 112 | }else{ |
| 113 | CHECK_HALF_MV(1, 1, mx-1, my ) |
| 114 | } |
| 115 | CHECK_HALF_MV(1, 0, mx-1, my ) |
| 116 | }else{ |
| 117 | CHECK_HALF_MV(1, 1, mx , my-1) |
| 118 | if(t+l<=b+r){ |
| 119 | CHECK_HALF_MV(1, 1, mx-1, my-1) |
| 120 | }else{ |
| 121 | CHECK_HALF_MV(1, 1, mx , my ) |
| 122 | } |
| 123 | CHECK_HALF_MV(1, 0, mx , my ) |
| 124 | } |
| 125 | }else{ |
| 126 | if(l<=r){ |
| 127 | if(t+l<=b+r){ |
| 128 | CHECK_HALF_MV(1, 1, mx-1, my-1) |
| 129 | }else{ |
| 130 | CHECK_HALF_MV(1, 1, mx , my ) |
| 131 | } |
| 132 | CHECK_HALF_MV(1, 0, mx-1, my) |
| 133 | CHECK_HALF_MV(1, 1, mx-1, my) |
| 134 | }else{ |
| 135 | if(t+r<=b+l){ |
| 136 | CHECK_HALF_MV(1, 1, mx , my-1) |
| 137 | }else{ |
| 138 | CHECK_HALF_MV(1, 1, mx-1, my) |
| 139 | } |
| 140 | CHECK_HALF_MV(1, 0, mx , my) |
| 141 | CHECK_HALF_MV(1, 1, mx , my) |
| 142 | } |
| 143 | CHECK_HALF_MV(0, 1, mx , my) |
| 144 | } |
| 145 | av_assert2(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2); |
| 146 | } |
| 147 | |
| 148 | *mx_ptr = bx; |
| 149 | *my_ptr = by; |
| 150 | |
| 151 | return dmin; |
| 152 | } |
| 153 | |
| 154 | static int no_sub_motion_search(MpegEncContext * s, |
| 155 | int *mx_ptr, int *my_ptr, int dmin, |
| 156 | int src_index, int ref_index, |
| 157 | int size, int h) |
| 158 | { |
| 159 | (*mx_ptr)<<=1; |
| 160 | (*my_ptr)<<=1; |
| 161 | return dmin; |
| 162 | } |
| 163 | |
| 164 | static inline int get_mb_score(MpegEncContext *s, int mx, int my, |
| 165 | int src_index, int ref_index, int size, |
| 166 | int h, int add_rate) |
| 167 | { |
| 168 | MotionEstContext * const c= &s->me; |
| 169 | const int penalty_factor= c->mb_penalty_factor; |
| 170 | const int flags= c->mb_flags; |
| 171 | const int qpel= flags & FLAG_QPEL; |
| 172 | const int mask= 1+2*qpel; |
| 173 | me_cmp_func cmp_sub, chroma_cmp_sub; |
| 174 | int d; |
| 175 | |
| 176 | LOAD_COMMON |
| 177 | |
| 178 | //FIXME factorize |
| 179 | |
| 180 | cmp_sub = s->mecc.mb_cmp[size]; |
| 181 | chroma_cmp_sub = s->mecc.mb_cmp[size + 1]; |
| 182 | |
| 183 | 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); |
| 184 | //FIXME check cbp before adding penalty for (0,0) vector |
| 185 | if(add_rate && (mx || my || size>0)) |
| 186 | d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor; |
| 187 | |
| 188 | return d; |
| 189 | } |
| 190 | |
| 191 | int ff_get_mb_score(MpegEncContext *s, int mx, int my, int src_index, |
| 192 | int ref_index, int size, int h, int add_rate) |
| 193 | { |
| 194 | return get_mb_score(s, mx, my, src_index, ref_index, size, h, add_rate); |
| 195 | } |
| 196 | |
| 197 | #define CHECK_QUARTER_MV(dx, dy, x, y)\ |
| 198 | {\ |
| 199 | const int hx= 4*(x)+(dx);\ |
| 200 | const int hy= 4*(y)+(dy);\ |
| 201 | d= cmp_qpel(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ |
| 202 | d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\ |
| 203 | COPY3_IF_LT(dmin, d, bx, hx, by, hy)\ |
| 204 | } |
| 205 | |
| 206 | static int qpel_motion_search(MpegEncContext * s, |
| 207 | int *mx_ptr, int *my_ptr, int dmin, |
| 208 | int src_index, int ref_index, |
| 209 | int size, int h) |
| 210 | { |
| 211 | MotionEstContext * const c= &s->me; |
| 212 | const int mx = *mx_ptr; |
| 213 | const int my = *my_ptr; |
| 214 | const int penalty_factor= c->sub_penalty_factor; |
| 215 | const unsigned map_generation = c->map_generation; |
| 216 | const int subpel_quality= c->avctx->me_subpel_quality; |
| 217 | uint32_t *map= c->map; |
| 218 | me_cmp_func cmpf, chroma_cmpf; |
| 219 | me_cmp_func cmp_sub, chroma_cmp_sub; |
| 220 | |
| 221 | LOAD_COMMON |
| 222 | int flags= c->sub_flags; |
| 223 | |
| 224 | cmpf = s->mecc.me_cmp[size]; |
| 225 | chroma_cmpf = s->mecc.me_cmp[size + 1]; // FIXME: factorize |
| 226 | //FIXME factorize |
| 227 | |
| 228 | cmp_sub = s->mecc.me_sub_cmp[size]; |
| 229 | chroma_cmp_sub = s->mecc.me_sub_cmp[size + 1]; |
| 230 | |
| 231 | if(c->skip){ //FIXME somehow move up (benchmark) |
| 232 | *mx_ptr = 0; |
| 233 | *my_ptr = 0; |
| 234 | return dmin; |
| 235 | } |
| 236 | |
| 237 | if(c->avctx->me_cmp != c->avctx->me_sub_cmp){ |
| 238 | dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags); |
| 239 | if(mx || my || size>0) |
| 240 | dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor; |
| 241 | } |
| 242 | |
| 243 | if (mx > xmin && mx < xmax && |
| 244 | my > ymin && my < ymax) { |
| 245 | int bx=4*mx, by=4*my; |
| 246 | int d= dmin; |
| 247 | int i, nx, ny; |
| 248 | const int index= (my<<ME_MAP_SHIFT) + mx; |
| 249 | const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)]; |
| 250 | const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]; |
| 251 | const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]; |
| 252 | const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)]; |
| 253 | const int c= score_map[(index )&(ME_MAP_SIZE-1)]; |
| 254 | int best[8]; |
| 255 | int best_pos[8][2]; |
| 256 | |
| 257 | memset(best, 64, sizeof(int)*8); |
| 258 | if(s->me.dia_size>=2){ |
| 259 | const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; |
| 260 | const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; |
| 261 | const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; |
| 262 | const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; |
| 263 | |
| 264 | for(ny= -3; ny <= 3; ny++){ |
| 265 | for(nx= -3; nx <= 3; nx++){ |
| 266 | //FIXME this could overflow (unlikely though) |
| 267 | const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t; |
| 268 | const int64_t c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c; |
| 269 | const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b; |
| 270 | int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10; |
| 271 | int i; |
| 272 | |
| 273 | if((nx&3)==0 && (ny&3)==0) continue; |
| 274 | |
| 275 | score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor; |
| 276 | |
| 277 | // if(nx&1) score-=1024*c->penalty_factor; |
| 278 | // if(ny&1) score-=1024*c->penalty_factor; |
| 279 | |
| 280 | for(i=0; i<8; i++){ |
| 281 | if(score < best[i]){ |
| 282 | memmove(&best[i+1], &best[i], sizeof(int)*(7-i)); |
| 283 | memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i)); |
| 284 | best[i]= score; |
| 285 | best_pos[i][0]= nx + 4*mx; |
| 286 | best_pos[i][1]= ny + 4*my; |
| 287 | break; |
| 288 | } |
| 289 | } |
| 290 | } |
| 291 | } |
| 292 | }else{ |
| 293 | int tl; |
| 294 | //FIXME this could overflow (unlikely though) |
| 295 | const int cx = 4*(r - l); |
| 296 | const int cx2= r + l - 2*c; |
| 297 | const int cy = 4*(b - t); |
| 298 | const int cy2= b + t - 2*c; |
| 299 | int cxy; |
| 300 | |
| 301 | if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME |
| 302 | tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; |
| 303 | }else{ |
| 304 | 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 |
| 305 | } |
| 306 | |
| 307 | cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c; |
| 308 | |
| 309 | av_assert2(16*cx2 + 4*cx + 32*c == 32*r); |
| 310 | av_assert2(16*cx2 - 4*cx + 32*c == 32*l); |
| 311 | av_assert2(16*cy2 + 4*cy + 32*c == 32*b); |
| 312 | av_assert2(16*cy2 - 4*cy + 32*c == 32*t); |
| 313 | av_assert2(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl); |
| 314 | |
| 315 | for(ny= -3; ny <= 3; ny++){ |
| 316 | for(nx= -3; nx <= 3; nx++){ |
| 317 | //FIXME this could overflow (unlikely though) |
| 318 | int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor |
| 319 | int i; |
| 320 | |
| 321 | if((nx&3)==0 && (ny&3)==0) continue; |
| 322 | |
| 323 | score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor; |
| 324 | // if(nx&1) score-=32*c->penalty_factor; |
| 325 | // if(ny&1) score-=32*c->penalty_factor; |
| 326 | |
| 327 | for(i=0; i<8; i++){ |
| 328 | if(score < best[i]){ |
| 329 | memmove(&best[i+1], &best[i], sizeof(int)*(7-i)); |
| 330 | memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i)); |
| 331 | best[i]= score; |
| 332 | best_pos[i][0]= nx + 4*mx; |
| 333 | best_pos[i][1]= ny + 4*my; |
| 334 | break; |
| 335 | } |
| 336 | } |
| 337 | } |
| 338 | } |
| 339 | } |
| 340 | for(i=0; i<subpel_quality; i++){ |
| 341 | nx= best_pos[i][0]; |
| 342 | ny= best_pos[i][1]; |
| 343 | CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2) |
| 344 | } |
| 345 | |
| 346 | av_assert2(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4); |
| 347 | |
| 348 | *mx_ptr = bx; |
| 349 | *my_ptr = by; |
| 350 | }else{ |
| 351 | *mx_ptr =4*mx; |
| 352 | *my_ptr =4*my; |
| 353 | } |
| 354 | |
| 355 | return dmin; |
| 356 | } |
| 357 | |
| 358 | |
| 359 | #define CHECK_MV(x,y)\ |
| 360 | {\ |
| 361 | const unsigned key = ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\ |
| 362 | const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\ |
| 363 | av_assert2((x) >= xmin);\ |
| 364 | av_assert2((x) <= xmax);\ |
| 365 | av_assert2((y) >= ymin);\ |
| 366 | av_assert2((y) <= ymax);\ |
| 367 | if(map[index]!=key){\ |
| 368 | d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ |
| 369 | map[index]= key;\ |
| 370 | score_map[index]= d;\ |
| 371 | d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\ |
| 372 | COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\ |
| 373 | }\ |
| 374 | } |
| 375 | |
| 376 | #define CHECK_CLIPPED_MV(ax,ay)\ |
| 377 | {\ |
| 378 | const int Lx= ax;\ |
| 379 | const int Ly= ay;\ |
| 380 | const int Lx2= FFMAX(xmin, FFMIN(Lx, xmax));\ |
| 381 | const int Ly2= FFMAX(ymin, FFMIN(Ly, ymax));\ |
| 382 | CHECK_MV(Lx2, Ly2)\ |
| 383 | } |
| 384 | |
| 385 | #define CHECK_MV_DIR(x,y,new_dir)\ |
| 386 | {\ |
| 387 | const unsigned key = ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\ |
| 388 | const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\ |
| 389 | if(map[index]!=key){\ |
| 390 | d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ |
| 391 | map[index]= key;\ |
| 392 | score_map[index]= d;\ |
| 393 | d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\ |
| 394 | if(d<dmin){\ |
| 395 | best[0]=x;\ |
| 396 | best[1]=y;\ |
| 397 | dmin=d;\ |
| 398 | next_dir= new_dir;\ |
| 399 | }\ |
| 400 | }\ |
| 401 | } |
| 402 | |
| 403 | #define check(x,y,S,v)\ |
| 404 | if( (x)<(xmin<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\ |
| 405 | if( (x)>(xmax<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\ |
| 406 | if( (y)<(ymin<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\ |
| 407 | if( (y)>(ymax<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\ |
| 408 | |
| 409 | #define LOAD_COMMON2\ |
| 410 | uint32_t *map= c->map;\ |
| 411 | const int qpel= flags&FLAG_QPEL;\ |
| 412 | const int shift= 1+qpel;\ |
| 413 | |
| 414 | static av_always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin, |
| 415 | int src_index, int ref_index, int const penalty_factor, |
| 416 | int size, int h, int flags) |
| 417 | { |
| 418 | MotionEstContext * const c= &s->me; |
| 419 | me_cmp_func cmpf, chroma_cmpf; |
| 420 | int next_dir=-1; |
| 421 | LOAD_COMMON |
| 422 | LOAD_COMMON2 |
| 423 | unsigned map_generation = c->map_generation; |
| 424 | |
| 425 | cmpf = s->mecc.me_cmp[size]; |
| 426 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 427 | |
| 428 | { /* ensure that the best point is in the MAP as h/qpel refinement needs it */ |
| 429 | const unsigned key = (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation; |
| 430 | const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1); |
| 431 | if(map[index]!=key){ //this will be executed only very rarey |
| 432 | score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags); |
| 433 | map[index]= key; |
| 434 | } |
| 435 | } |
| 436 | |
| 437 | for(;;){ |
| 438 | int d; |
| 439 | const int dir= next_dir; |
| 440 | const int x= best[0]; |
| 441 | const int y= best[1]; |
| 442 | next_dir=-1; |
| 443 | |
| 444 | if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0) |
| 445 | if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1) |
| 446 | if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2) |
| 447 | if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3) |
| 448 | |
| 449 | if(next_dir==-1){ |
| 450 | return dmin; |
| 451 | } |
| 452 | } |
| 453 | } |
| 454 | |
| 455 | static int funny_diamond_search(MpegEncContext * s, int *best, int dmin, |
| 456 | int src_index, int ref_index, int const penalty_factor, |
| 457 | int size, int h, int flags) |
| 458 | { |
| 459 | MotionEstContext * const c= &s->me; |
| 460 | me_cmp_func cmpf, chroma_cmpf; |
| 461 | int dia_size; |
| 462 | LOAD_COMMON |
| 463 | LOAD_COMMON2 |
| 464 | unsigned map_generation = c->map_generation; |
| 465 | |
| 466 | cmpf = s->mecc.me_cmp[size]; |
| 467 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 468 | |
| 469 | for(dia_size=1; dia_size<=4; dia_size++){ |
| 470 | int dir; |
| 471 | const int x= best[0]; |
| 472 | const int y= best[1]; |
| 473 | |
| 474 | if(dia_size&(dia_size-1)) continue; |
| 475 | |
| 476 | if( x + dia_size > xmax |
| 477 | || x - dia_size < xmin |
| 478 | || y + dia_size > ymax |
| 479 | || y - dia_size < ymin) |
| 480 | continue; |
| 481 | |
| 482 | for(dir= 0; dir<dia_size; dir+=2){ |
| 483 | int d; |
| 484 | |
| 485 | CHECK_MV(x + dir , y + dia_size - dir); |
| 486 | CHECK_MV(x + dia_size - dir, y - dir ); |
| 487 | CHECK_MV(x - dir , y - dia_size + dir); |
| 488 | CHECK_MV(x - dia_size + dir, y + dir ); |
| 489 | } |
| 490 | |
| 491 | if(x!=best[0] || y!=best[1]) |
| 492 | dia_size=0; |
| 493 | } |
| 494 | return dmin; |
| 495 | } |
| 496 | |
| 497 | static int hex_search(MpegEncContext * s, int *best, int dmin, |
| 498 | int src_index, int ref_index, int const penalty_factor, |
| 499 | int size, int h, int flags, int dia_size) |
| 500 | { |
| 501 | MotionEstContext * const c= &s->me; |
| 502 | me_cmp_func cmpf, chroma_cmpf; |
| 503 | LOAD_COMMON |
| 504 | LOAD_COMMON2 |
| 505 | unsigned map_generation = c->map_generation; |
| 506 | int x,y,d; |
| 507 | const int dec= dia_size & (dia_size-1); |
| 508 | |
| 509 | cmpf = s->mecc.me_cmp[size]; |
| 510 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 511 | |
| 512 | for(;dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){ |
| 513 | do{ |
| 514 | x= best[0]; |
| 515 | y= best[1]; |
| 516 | |
| 517 | CHECK_CLIPPED_MV(x -dia_size , y); |
| 518 | CHECK_CLIPPED_MV(x+ dia_size , y); |
| 519 | CHECK_CLIPPED_MV(x+( dia_size>>1), y+dia_size); |
| 520 | CHECK_CLIPPED_MV(x+( dia_size>>1), y-dia_size); |
| 521 | if(dia_size>1){ |
| 522 | CHECK_CLIPPED_MV(x+(-dia_size>>1), y+dia_size); |
| 523 | CHECK_CLIPPED_MV(x+(-dia_size>>1), y-dia_size); |
| 524 | } |
| 525 | }while(best[0] != x || best[1] != y); |
| 526 | } |
| 527 | |
| 528 | return dmin; |
| 529 | } |
| 530 | |
| 531 | static int l2s_dia_search(MpegEncContext * s, int *best, int dmin, |
| 532 | int src_index, int ref_index, int const penalty_factor, |
| 533 | int size, int h, int flags) |
| 534 | { |
| 535 | MotionEstContext * const c= &s->me; |
| 536 | me_cmp_func cmpf, chroma_cmpf; |
| 537 | LOAD_COMMON |
| 538 | LOAD_COMMON2 |
| 539 | unsigned map_generation = c->map_generation; |
| 540 | int x,y,i,d; |
| 541 | int dia_size= c->dia_size&0xFF; |
| 542 | const int dec= dia_size & (dia_size-1); |
| 543 | static const int hex[8][2]={{-2, 0}, {-1,-1}, { 0,-2}, { 1,-1}, |
| 544 | { 2, 0}, { 1, 1}, { 0, 2}, {-1, 1}}; |
| 545 | |
| 546 | cmpf = s->mecc.me_cmp[size]; |
| 547 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 548 | |
| 549 | for(; dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){ |
| 550 | do{ |
| 551 | x= best[0]; |
| 552 | y= best[1]; |
| 553 | for(i=0; i<8; i++){ |
| 554 | CHECK_CLIPPED_MV(x+hex[i][0]*dia_size, y+hex[i][1]*dia_size); |
| 555 | } |
| 556 | }while(best[0] != x || best[1] != y); |
| 557 | } |
| 558 | |
| 559 | x= best[0]; |
| 560 | y= best[1]; |
| 561 | CHECK_CLIPPED_MV(x+1, y); |
| 562 | CHECK_CLIPPED_MV(x, y+1); |
| 563 | CHECK_CLIPPED_MV(x-1, y); |
| 564 | CHECK_CLIPPED_MV(x, y-1); |
| 565 | |
| 566 | return dmin; |
| 567 | } |
| 568 | |
| 569 | static int umh_search(MpegEncContext * s, int *best, int dmin, |
| 570 | int src_index, int ref_index, int const penalty_factor, |
| 571 | int size, int h, int flags) |
| 572 | { |
| 573 | MotionEstContext * const c= &s->me; |
| 574 | me_cmp_func cmpf, chroma_cmpf; |
| 575 | LOAD_COMMON |
| 576 | LOAD_COMMON2 |
| 577 | unsigned map_generation = c->map_generation; |
| 578 | int x,y,x2,y2, i, j, d; |
| 579 | const int dia_size= c->dia_size&0xFE; |
| 580 | static const int hex[16][2]={{-4,-2}, {-4,-1}, {-4, 0}, {-4, 1}, {-4, 2}, |
| 581 | { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2}, |
| 582 | {-2, 3}, { 0, 4}, { 2, 3}, |
| 583 | {-2,-3}, { 0,-4}, { 2,-3},}; |
| 584 | |
| 585 | cmpf = s->mecc.me_cmp[size]; |
| 586 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 587 | |
| 588 | x= best[0]; |
| 589 | y= best[1]; |
| 590 | for(x2=FFMAX(x-dia_size+1, xmin); x2<=FFMIN(x+dia_size-1,xmax); x2+=2){ |
| 591 | CHECK_MV(x2, y); |
| 592 | } |
| 593 | for(y2=FFMAX(y-dia_size/2+1, ymin); y2<=FFMIN(y+dia_size/2-1,ymax); y2+=2){ |
| 594 | CHECK_MV(x, y2); |
| 595 | } |
| 596 | |
| 597 | x= best[0]; |
| 598 | y= best[1]; |
| 599 | for(y2=FFMAX(y-2, ymin); y2<=FFMIN(y+2,ymax); y2++){ |
| 600 | for(x2=FFMAX(x-2, xmin); x2<=FFMIN(x+2,xmax); x2++){ |
| 601 | CHECK_MV(x2, y2); |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | //FIXME prevent the CLIP stuff |
| 606 | |
| 607 | for(j=1; j<=dia_size/4; j++){ |
| 608 | for(i=0; i<16; i++){ |
| 609 | CHECK_CLIPPED_MV(x+hex[i][0]*j, y+hex[i][1]*j); |
| 610 | } |
| 611 | } |
| 612 | |
| 613 | return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, 2); |
| 614 | } |
| 615 | |
| 616 | static int full_search(MpegEncContext * s, int *best, int dmin, |
| 617 | int src_index, int ref_index, int const penalty_factor, |
| 618 | int size, int h, int flags) |
| 619 | { |
| 620 | MotionEstContext * const c= &s->me; |
| 621 | me_cmp_func cmpf, chroma_cmpf; |
| 622 | LOAD_COMMON |
| 623 | LOAD_COMMON2 |
| 624 | unsigned map_generation = c->map_generation; |
| 625 | int x,y, d; |
| 626 | const int dia_size= c->dia_size&0xFF; |
| 627 | |
| 628 | cmpf = s->mecc.me_cmp[size]; |
| 629 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 630 | |
| 631 | for(y=FFMAX(-dia_size, ymin); y<=FFMIN(dia_size,ymax); y++){ |
| 632 | for(x=FFMAX(-dia_size, xmin); x<=FFMIN(dia_size,xmax); x++){ |
| 633 | CHECK_MV(x, y); |
| 634 | } |
| 635 | } |
| 636 | |
| 637 | x= best[0]; |
| 638 | y= best[1]; |
| 639 | d= dmin; |
| 640 | CHECK_CLIPPED_MV(x , y); |
| 641 | CHECK_CLIPPED_MV(x+1, y); |
| 642 | CHECK_CLIPPED_MV(x, y+1); |
| 643 | CHECK_CLIPPED_MV(x-1, y); |
| 644 | CHECK_CLIPPED_MV(x, y-1); |
| 645 | best[0]= x; |
| 646 | best[1]= y; |
| 647 | |
| 648 | return d; |
| 649 | } |
| 650 | |
| 651 | #define SAB_CHECK_MV(ax,ay)\ |
| 652 | {\ |
| 653 | const unsigned key = ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\ |
| 654 | const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\ |
| 655 | if(map[index]!=key){\ |
| 656 | d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ |
| 657 | map[index]= key;\ |
| 658 | score_map[index]= d;\ |
| 659 | d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\ |
| 660 | if(d < minima[minima_count-1].height){\ |
| 661 | int j=0;\ |
| 662 | \ |
| 663 | while(d >= minima[j].height) j++;\ |
| 664 | \ |
| 665 | memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\ |
| 666 | \ |
| 667 | minima[j].checked= 0;\ |
| 668 | minima[j].height= d;\ |
| 669 | minima[j].x= ax;\ |
| 670 | minima[j].y= ay;\ |
| 671 | \ |
| 672 | i=-1;\ |
| 673 | continue;\ |
| 674 | }\ |
| 675 | }\ |
| 676 | } |
| 677 | |
| 678 | #define MAX_SAB_SIZE ME_MAP_SIZE |
| 679 | static int sab_diamond_search(MpegEncContext * s, int *best, int dmin, |
| 680 | int src_index, int ref_index, int const penalty_factor, |
| 681 | int size, int h, int flags) |
| 682 | { |
| 683 | MotionEstContext * const c= &s->me; |
| 684 | me_cmp_func cmpf, chroma_cmpf; |
| 685 | Minima minima[MAX_SAB_SIZE]; |
| 686 | const int minima_count= FFABS(c->dia_size); |
| 687 | int i, j; |
| 688 | LOAD_COMMON |
| 689 | LOAD_COMMON2 |
| 690 | unsigned map_generation = c->map_generation; |
| 691 | |
| 692 | av_assert1(minima_count <= MAX_SAB_SIZE); |
| 693 | |
| 694 | cmpf = s->mecc.me_cmp[size]; |
| 695 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 696 | |
| 697 | /*Note j<MAX_SAB_SIZE is needed if MAX_SAB_SIZE < ME_MAP_SIZE as j can |
| 698 | become larger due to MVs overflowing their ME_MAP_MV_BITS bits space in map |
| 699 | */ |
| 700 | for(j=i=0; i<ME_MAP_SIZE && j<MAX_SAB_SIZE; i++){ |
| 701 | uint32_t key= map[i]; |
| 702 | |
| 703 | key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1)); |
| 704 | |
| 705 | if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue; |
| 706 | |
| 707 | minima[j].height= score_map[i]; |
| 708 | minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS; |
| 709 | minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1); |
| 710 | minima[j].x-= (1<<(ME_MAP_MV_BITS-1)); |
| 711 | minima[j].y-= (1<<(ME_MAP_MV_BITS-1)); |
| 712 | |
| 713 | // all entries in map should be in range except if the mv overflows their ME_MAP_MV_BITS bits space |
| 714 | if( minima[j].x > xmax || minima[j].x < xmin |
| 715 | || minima[j].y > ymax || minima[j].y < ymin) |
| 716 | continue; |
| 717 | |
| 718 | minima[j].checked=0; |
| 719 | if(minima[j].x || minima[j].y) |
| 720 | minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor; |
| 721 | |
| 722 | j++; |
| 723 | } |
| 724 | |
| 725 | qsort(minima, j, sizeof(Minima), minima_cmp); |
| 726 | |
| 727 | for(; j<minima_count; j++){ |
| 728 | minima[j].height=256*256*256*64; |
| 729 | minima[j].checked=0; |
| 730 | minima[j].x= minima[j].y=0; |
| 731 | } |
| 732 | |
| 733 | for(i=0; i<minima_count; i++){ |
| 734 | const int x= minima[i].x; |
| 735 | const int y= minima[i].y; |
| 736 | int d; |
| 737 | |
| 738 | if(minima[i].checked) continue; |
| 739 | |
| 740 | if( x >= xmax || x <= xmin |
| 741 | || y >= ymax || y <= ymin) |
| 742 | continue; |
| 743 | |
| 744 | SAB_CHECK_MV(x-1, y) |
| 745 | SAB_CHECK_MV(x+1, y) |
| 746 | SAB_CHECK_MV(x , y-1) |
| 747 | SAB_CHECK_MV(x , y+1) |
| 748 | |
| 749 | minima[i].checked= 1; |
| 750 | } |
| 751 | |
| 752 | best[0]= minima[0].x; |
| 753 | best[1]= minima[0].y; |
| 754 | dmin= minima[0].height; |
| 755 | |
| 756 | if( best[0] < xmax && best[0] > xmin |
| 757 | && best[1] < ymax && best[1] > ymin){ |
| 758 | int d; |
| 759 | //ensure that the refernece samples for hpel refinement are in the map |
| 760 | CHECK_MV(best[0]-1, best[1]) |
| 761 | CHECK_MV(best[0]+1, best[1]) |
| 762 | CHECK_MV(best[0], best[1]-1) |
| 763 | CHECK_MV(best[0], best[1]+1) |
| 764 | } |
| 765 | return dmin; |
| 766 | } |
| 767 | |
| 768 | static int var_diamond_search(MpegEncContext * s, int *best, int dmin, |
| 769 | int src_index, int ref_index, int const penalty_factor, |
| 770 | int size, int h, int flags) |
| 771 | { |
| 772 | MotionEstContext * const c= &s->me; |
| 773 | me_cmp_func cmpf, chroma_cmpf; |
| 774 | int dia_size; |
| 775 | LOAD_COMMON |
| 776 | LOAD_COMMON2 |
| 777 | unsigned map_generation = c->map_generation; |
| 778 | |
| 779 | cmpf = s->mecc.me_cmp[size]; |
| 780 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 781 | |
| 782 | for(dia_size=1; dia_size<=c->dia_size; dia_size++){ |
| 783 | int dir, start, end; |
| 784 | const int x= best[0]; |
| 785 | const int y= best[1]; |
| 786 | |
| 787 | start= FFMAX(0, y + dia_size - ymax); |
| 788 | end = FFMIN(dia_size, xmax - x + 1); |
| 789 | for(dir= start; dir<end; dir++){ |
| 790 | int d; |
| 791 | |
| 792 | //check(x + dir,y + dia_size - dir,0, a0) |
| 793 | CHECK_MV(x + dir , y + dia_size - dir); |
| 794 | } |
| 795 | |
| 796 | start= FFMAX(0, x + dia_size - xmax); |
| 797 | end = FFMIN(dia_size, y - ymin + 1); |
| 798 | for(dir= start; dir<end; dir++){ |
| 799 | int d; |
| 800 | |
| 801 | //check(x + dia_size - dir, y - dir,0, a1) |
| 802 | CHECK_MV(x + dia_size - dir, y - dir ); |
| 803 | } |
| 804 | |
| 805 | start= FFMAX(0, -y + dia_size + ymin ); |
| 806 | end = FFMIN(dia_size, x - xmin + 1); |
| 807 | for(dir= start; dir<end; dir++){ |
| 808 | int d; |
| 809 | |
| 810 | //check(x - dir,y - dia_size + dir,0, a2) |
| 811 | CHECK_MV(x - dir , y - dia_size + dir); |
| 812 | } |
| 813 | |
| 814 | start= FFMAX(0, -x + dia_size + xmin ); |
| 815 | end = FFMIN(dia_size, ymax - y + 1); |
| 816 | for(dir= start; dir<end; dir++){ |
| 817 | int d; |
| 818 | |
| 819 | //check(x - dia_size + dir, y + dir,0, a3) |
| 820 | CHECK_MV(x - dia_size + dir, y + dir ); |
| 821 | } |
| 822 | |
| 823 | if(x!=best[0] || y!=best[1]) |
| 824 | dia_size=0; |
| 825 | } |
| 826 | return dmin; |
| 827 | } |
| 828 | |
| 829 | static av_always_inline int diamond_search(MpegEncContext * s, int *best, int dmin, |
| 830 | int src_index, int ref_index, int const penalty_factor, |
| 831 | int size, int h, int flags){ |
| 832 | MotionEstContext * const c= &s->me; |
| 833 | if(c->dia_size==-1) |
| 834 | return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 835 | else if(c->dia_size<-1) |
| 836 | return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 837 | else if(c->dia_size<2) |
| 838 | return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 839 | else if(c->dia_size>1024) |
| 840 | return full_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 841 | else if(c->dia_size>768) |
| 842 | return umh_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 843 | else if(c->dia_size>512) |
| 844 | return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, c->dia_size&0xFF); |
| 845 | else if(c->dia_size>256) |
| 846 | return l2s_dia_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 847 | else |
| 848 | return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 849 | } |
| 850 | |
| 851 | /** |
| 852 | @param P a list of candidate mvs to check before starting the |
| 853 | iterative search. If one of the candidates is close to the optimal mv, then |
| 854 | it takes fewer iterations. And it increases the chance that we find the |
| 855 | optimal mv. |
| 856 | */ |
| 857 | static av_always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr, |
| 858 | int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2], |
| 859 | int ref_mv_scale, int flags, int size, int h) |
| 860 | { |
| 861 | MotionEstContext * const c= &s->me; |
| 862 | int best[2]={0, 0}; /**< x and y coordinates of the best motion vector. |
| 863 | i.e. the difference between the position of the |
| 864 | block currently being encoded and the position of |
| 865 | the block chosen to predict it from. */ |
| 866 | int d; ///< the score (cmp + penalty) of any given mv |
| 867 | int dmin; /**< the best value of d, i.e. the score |
| 868 | corresponding to the mv stored in best[]. */ |
| 869 | unsigned map_generation; |
| 870 | int penalty_factor; |
| 871 | const int ref_mv_stride= s->mb_stride; //pass as arg FIXME |
| 872 | const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME |
| 873 | me_cmp_func cmpf, chroma_cmpf; |
| 874 | |
| 875 | LOAD_COMMON |
| 876 | LOAD_COMMON2 |
| 877 | |
| 878 | if(c->pre_pass){ |
| 879 | penalty_factor= c->pre_penalty_factor; |
| 880 | cmpf = s->mecc.me_pre_cmp[size]; |
| 881 | chroma_cmpf = s->mecc.me_pre_cmp[size + 1]; |
| 882 | }else{ |
| 883 | penalty_factor= c->penalty_factor; |
| 884 | cmpf = s->mecc.me_cmp[size]; |
| 885 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 886 | } |
| 887 | |
| 888 | map_generation= update_map_generation(c); |
| 889 | |
| 890 | av_assert2(cmpf); |
| 891 | dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags); |
| 892 | map[0]= map_generation; |
| 893 | score_map[0]= dmin; |
| 894 | |
| 895 | //FIXME precalc first term below? |
| 896 | if ((s->pict_type == AV_PICTURE_TYPE_B && !(c->flags & FLAG_DIRECT)) || |
| 897 | s->mpv_flags & FF_MPV_FLAG_MV0) |
| 898 | dmin += (mv_penalty[pred_x] + mv_penalty[pred_y])*penalty_factor; |
| 899 | |
| 900 | /* first line */ |
| 901 | if (s->first_slice_line) { |
| 902 | CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) |
| 903 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
| 904 | (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) |
| 905 | }else{ |
| 906 | if(dmin<((h*h*s->avctx->mv0_threshold)>>8) |
| 907 | && ( P_LEFT[0] |P_LEFT[1] |
| 908 | |P_TOP[0] |P_TOP[1] |
| 909 | |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){ |
| 910 | *mx_ptr= 0; |
| 911 | *my_ptr= 0; |
| 912 | c->skip=1; |
| 913 | return dmin; |
| 914 | } |
| 915 | CHECK_MV( P_MEDIAN[0] >>shift , P_MEDIAN[1] >>shift) |
| 916 | CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)-1) |
| 917 | CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)+1) |
| 918 | CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)-1, (P_MEDIAN[1]>>shift) ) |
| 919 | CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)+1, (P_MEDIAN[1]>>shift) ) |
| 920 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
| 921 | (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) |
| 922 | CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift) |
| 923 | CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift) |
| 924 | CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) |
| 925 | } |
| 926 | if(dmin>h*h*4){ |
| 927 | if(c->pre_pass){ |
| 928 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16, |
| 929 | (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16) |
| 930 | if(!s->first_slice_line) |
| 931 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, |
| 932 | (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) |
| 933 | }else{ |
| 934 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, |
| 935 | (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16) |
| 936 | if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line |
| 937 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, |
| 938 | (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) |
| 939 | } |
| 940 | } |
| 941 | |
| 942 | if(c->avctx->last_predictor_count){ |
| 943 | const int count= c->avctx->last_predictor_count; |
| 944 | const int xstart= FFMAX(0, s->mb_x - count); |
| 945 | const int ystart= FFMAX(0, s->mb_y - count); |
| 946 | const int xend= FFMIN(s->mb_width , s->mb_x + count + 1); |
| 947 | const int yend= FFMIN(s->mb_height, s->mb_y + count + 1); |
| 948 | int mb_y; |
| 949 | |
| 950 | for(mb_y=ystart; mb_y<yend; mb_y++){ |
| 951 | int mb_x; |
| 952 | for(mb_x=xstart; mb_x<xend; mb_x++){ |
| 953 | const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride; |
| 954 | int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16; |
| 955 | int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16; |
| 956 | |
| 957 | if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue; |
| 958 | CHECK_MV(mx,my) |
| 959 | } |
| 960 | } |
| 961 | } |
| 962 | |
| 963 | //check(best[0],best[1],0, b0) |
| 964 | dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 965 | |
| 966 | //check(best[0],best[1],0, b1) |
| 967 | *mx_ptr= best[0]; |
| 968 | *my_ptr= best[1]; |
| 969 | |
| 970 | return dmin; |
| 971 | } |
| 972 | |
| 973 | //this function is dedicated to the braindamaged gcc |
| 974 | int ff_epzs_motion_search(MpegEncContext *s, int *mx_ptr, int *my_ptr, |
| 975 | int P[10][2], int src_index, int ref_index, |
| 976 | int16_t (*last_mv)[2], int ref_mv_scale, |
| 977 | int size, int h) |
| 978 | { |
| 979 | MotionEstContext * const c= &s->me; |
| 980 | //FIXME convert other functions in the same way if faster |
| 981 | if(c->flags==0 && h==16 && size==0){ |
| 982 | return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16); |
| 983 | // case FLAG_QPEL: |
| 984 | // return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL); |
| 985 | }else{ |
| 986 | return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h); |
| 987 | } |
| 988 | } |
| 989 | |
| 990 | static int epzs_motion_search4(MpegEncContext * s, |
| 991 | int *mx_ptr, int *my_ptr, int P[10][2], |
| 992 | int src_index, int ref_index, int16_t (*last_mv)[2], |
| 993 | int ref_mv_scale) |
| 994 | { |
| 995 | MotionEstContext * const c= &s->me; |
| 996 | int best[2]={0, 0}; |
| 997 | int d, dmin; |
| 998 | unsigned map_generation; |
| 999 | const int penalty_factor= c->penalty_factor; |
| 1000 | const int size=1; |
| 1001 | const int h=8; |
| 1002 | const int ref_mv_stride= s->mb_stride; |
| 1003 | const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride; |
| 1004 | me_cmp_func cmpf, chroma_cmpf; |
| 1005 | LOAD_COMMON |
| 1006 | int flags= c->flags; |
| 1007 | LOAD_COMMON2 |
| 1008 | |
| 1009 | cmpf = s->mecc.me_cmp[size]; |
| 1010 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 1011 | |
| 1012 | map_generation= update_map_generation(c); |
| 1013 | |
| 1014 | dmin = 1000000; |
| 1015 | |
| 1016 | /* first line */ |
| 1017 | if (s->first_slice_line) { |
| 1018 | CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) |
| 1019 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
| 1020 | (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) |
| 1021 | CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) |
| 1022 | }else{ |
| 1023 | CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) |
| 1024 | //FIXME try some early stop |
| 1025 | CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift) |
| 1026 | CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) |
| 1027 | CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift) |
| 1028 | CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) |
| 1029 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
| 1030 | (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) |
| 1031 | } |
| 1032 | if(dmin>64*4){ |
| 1033 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, |
| 1034 | (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16) |
| 1035 | if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line |
| 1036 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, |
| 1037 | (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) |
| 1038 | } |
| 1039 | |
| 1040 | dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 1041 | |
| 1042 | *mx_ptr= best[0]; |
| 1043 | *my_ptr= best[1]; |
| 1044 | |
| 1045 | return dmin; |
| 1046 | } |
| 1047 | |
| 1048 | //try to merge with above FIXME (needs PSNR test) |
| 1049 | static int epzs_motion_search2(MpegEncContext * s, |
| 1050 | int *mx_ptr, int *my_ptr, int P[10][2], |
| 1051 | int src_index, int ref_index, int16_t (*last_mv)[2], |
| 1052 | int ref_mv_scale) |
| 1053 | { |
| 1054 | MotionEstContext * const c= &s->me; |
| 1055 | int best[2]={0, 0}; |
| 1056 | int d, dmin; |
| 1057 | unsigned map_generation; |
| 1058 | const int penalty_factor= c->penalty_factor; |
| 1059 | const int size=0; //FIXME pass as arg |
| 1060 | const int h=8; |
| 1061 | const int ref_mv_stride= s->mb_stride; |
| 1062 | const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride; |
| 1063 | me_cmp_func cmpf, chroma_cmpf; |
| 1064 | LOAD_COMMON |
| 1065 | int flags= c->flags; |
| 1066 | LOAD_COMMON2 |
| 1067 | |
| 1068 | cmpf = s->mecc.me_cmp[size]; |
| 1069 | chroma_cmpf = s->mecc.me_cmp[size + 1]; |
| 1070 | |
| 1071 | map_generation= update_map_generation(c); |
| 1072 | |
| 1073 | dmin = 1000000; |
| 1074 | |
| 1075 | /* first line */ |
| 1076 | if (s->first_slice_line) { |
| 1077 | CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) |
| 1078 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
| 1079 | (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) |
| 1080 | CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) |
| 1081 | }else{ |
| 1082 | CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) |
| 1083 | //FIXME try some early stop |
| 1084 | CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift) |
| 1085 | CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) |
| 1086 | CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift) |
| 1087 | CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) |
| 1088 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
| 1089 | (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) |
| 1090 | } |
| 1091 | if(dmin>64*4){ |
| 1092 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, |
| 1093 | (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16) |
| 1094 | if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line |
| 1095 | CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, |
| 1096 | (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) |
| 1097 | } |
| 1098 | |
| 1099 | dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); |
| 1100 | |
| 1101 | *mx_ptr= best[0]; |
| 1102 | *my_ptr= best[1]; |
| 1103 | |
| 1104 | return dmin; |
| 1105 | } |