1 /* ------------------------------------------------------------------ 2 * Copyright (C) 1998-2009 PacketVideo 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either 13 * express or implied. 14 * See the License for the specific language governing permissions 15 * and limitations under the License. 16 * ------------------------------------------------------------------- 17 */ 18 #include "mp4def.h" 19 #include "mp4enc_lib.h" 20 #include "mp4lib_int.h" 21 #include "m4venc_oscl.h" 22 23 //#define PRINT_MV 24 #define MIN_GOP 1 /* minimum size of GOP, 1/23/01, need to be tested */ 25 26 #define CANDIDATE_DISTANCE 0 /* distance candidate from one another to consider as a distinct one */ 27 /* shouldn't be more than 3 */ 28 29 #define ZERO_MV_PREF 0 /* 0: bias (0,0)MV before full-pel search, lowest complexity*/ 30 /* 1: bias (0,0)MV after full-pel search, before half-pel, highest comp */ 31 /* 2: bias (0,0)MV after half-pel, high comp, better PSNR */ 32 33 #define RASTER_REFRESH /* instead of random INTRA refresh, do raster scan, 2/26/01 */ 34 35 #ifdef RASTER_REFRESH 36 #define TARGET_REFRESH_PER_REGION 4 /* , no. MB per frame to be INTRA refreshed */ 37 #else 38 #define TARGET_REFRESH_PER_REGION 1 /* , no. MB per region to be INTRA refreshed */ 39 #endif 40 41 #define ALL_CAND_EQUAL 10 /* any number greater than 5 will work */ 42 43 #define NumPixelMB 256 /* number of pixels used in SAD calculation */ 44 45 #define DEF_8X8_WIN 3 /* search region for 8x8 MVs around the 16x16 MV */ 46 #define MB_Nb 256 47 48 #define PREF_NULL_VEC 129 /* for zero vector bias */ 49 #define PREF_16_VEC 129 /* 1MV bias versus 4MVs*/ 50 #define PREF_INTRA 512 /* bias for INTRA coding */ 51 52 const static Int tab_exclude[9][9] = // [last_loc][curr_loc] 53 { 54 {0, 0, 0, 0, 0, 0, 0, 0, 0}, 55 {0, 0, 0, 0, 1, 1, 1, 0, 0}, 56 {0, 0, 0, 0, 1, 1, 1, 1, 1}, 57 {0, 0, 0, 0, 0, 0, 1, 1, 1}, 58 {0, 1, 1, 0, 0, 0, 1, 1, 1}, 59 {0, 1, 1, 0, 0, 0, 0, 0, 1}, 60 {0, 1, 1, 1, 1, 0, 0, 0, 1}, 61 {0, 0, 1, 1, 1, 0, 0, 0, 0}, 62 {0, 0, 1, 1, 1, 1, 1, 0, 0} 63 }; //to decide whether to continue or compute 64 65 const static Int refine_next[8][2] = /* [curr_k][increment] */ 66 { 67 {0, 0}, {2, 0}, {1, 1}, {0, 2}, { -1, 1}, { -2, 0}, { -1, -1}, {0, -2} 68 }; 69 70 #ifdef __cplusplus 71 extern "C" 72 { 73 #endif 74 75 void MBMotionSearch(VideoEncData *video, UChar *cur, UChar *best_cand[], 76 Int i0, Int j0, Int type_pred, Int fullsearch, Int *hp_guess); 77 78 Int fullsearch(VideoEncData *video, Vol *currVol, UChar *ref, UChar *cur, 79 Int *imin, Int *jmin, Int ilow, Int ihigh, Int jlow, Int jhigh); 80 Int fullsearchBlk(VideoEncData *video, Vol *currVol, UChar *cent, UChar *cur, 81 Int *imin, Int *jmin, Int ilow, Int ihigh, Int jlow, Int jhigh, Int range); 82 void CandidateSelection(Int *mvx, Int *mvy, Int *num_can, Int imb, Int jmb, 83 VideoEncData *video, Int type_pred); 84 void RasterIntraUpdate(UChar *intraArray, UChar *Mode, Int totalMB, Int numRefresh); 85 void ResetIntraUpdate(UChar *intraArray, Int totalMB); 86 void ResetIntraUpdateRegion(UChar *intraArray, Int start_i, Int rwidth, 87 Int start_j, Int rheight, Int mbwidth, Int mbheight); 88 89 void MoveNeighborSAD(Int dn[], Int new_loc); 90 Int FindMin(Int dn[]); 91 void PrepareCurMB(VideoEncData *video, UChar *cur); 92 93 #ifdef __cplusplus 94 } 95 #endif 96 97 /***************************************/ 98 /* 2/28/01, for HYPOTHESIS TESTING */ 99 #ifdef HTFM /* defined in mp4def.h */ 100 #ifdef __cplusplus 101 extern "C" 102 { 103 #endif 104 void CalcThreshold(double pf, double exp_lamda[], Int nrmlz_th[]); 105 void HTFMPrepareCurMB(VideoEncData *video, HTFM_Stat *htfm_stat, UChar *cur); 106 #ifdef __cplusplus 107 } 108 #endif 109 110 111 #define HTFM_Pf 0.25 /* 3/2/1, probability of false alarm, can be varied from 0 to 0.5 */ 112 /***************************************/ 113 #endif 114 115 #ifdef _SAD_STAT 116 ULong num_MB = 0; 117 ULong num_HP_MB = 0; 118 ULong num_Blk = 0; 119 ULong num_HP_Blk = 0; 120 ULong num_cand = 0; 121 ULong num_better_hp = 0; 122 ULong i_dist_from_guess = 0; 123 ULong j_dist_from_guess = 0; 124 ULong num_hp_not_zero = 0; 125 #endif 126 127 128 129 /*================================================================== 130 Function: MotionEstimation 131 Date: 10/3/2000 132 Purpose: Go through all macroblock for motion search and 133 determine scene change detection. 134 ====================================================================*/ 135 136 void MotionEstimation(VideoEncData *video) 137 { 138 UChar use_4mv = video->encParams->MV8x8_Enabled; 139 Vol *currVol = video->vol[video->currLayer]; 140 Vop *currVop = video->currVop; 141 VideoEncFrameIO *currFrame = video->input; 142 Int i, j, comp; 143 Int mbwidth = currVol->nMBPerRow; 144 Int mbheight = currVol->nMBPerCol; 145 Int totalMB = currVol->nTotalMB; 146 Int width = currFrame->pitch; 147 UChar *mode_mb, *Mode = video->headerInfo.Mode; 148 MOT *mot_mb, **mot = video->mot; 149 UChar *intraArray = video->intraArray; 150 Int FS_en = video->encParams->FullSearch_Enabled; 151 void (*ComputeMBSum)(UChar *, Int, MOT *) = video->functionPointer->ComputeMBSum; 152 void (*ChooseMode)(UChar*, UChar*, Int, Int) = video->functionPointer->ChooseMode; 153 154 Int numIntra, start_i, numLoop, incr_i; 155 Int mbnum, offset; 156 UChar *cur, *best_cand[5]; 157 Int sad8 = 0, sad16 = 0; 158 Int totalSAD = 0; /* average SAD for rate control */ 159 Int skip_halfpel_4mv; 160 Int f_code_p, f_code_n, max_mag = 0, min_mag = 0; 161 Int type_pred; 162 Int xh[5] = {0, 0, 0, 0, 0}; 163 Int yh[5] = {0, 0, 0, 0, 0}; /* half-pel */ 164 UChar hp_mem4MV[17*17*4]; 165 166 #ifdef HTFM 167 /***** HYPOTHESIS TESTING ********/ /* 2/28/01 */ 168 Int collect = 0; 169 HTFM_Stat htfm_stat; 170 double newvar[16]; 171 double exp_lamda[15]; 172 /*********************************/ 173 #endif 174 Int hp_guess = 0; 175 #ifdef PRINT_MV 176 FILE *fp_debug; 177 #endif 178 179 // FILE *fstat; 180 // static int frame_num = 0; 181 182 offset = 0; 183 184 if (video->currVop->predictionType == I_VOP) 185 { /* compute the SAV */ 186 mbnum = 0; 187 cur = currFrame->yChan; 188 189 for (j = 0; j < mbheight; j++) 190 { 191 for (i = 0; i < mbwidth; i++) 192 { 193 video->mbnum = mbnum; 194 mot_mb = mot[mbnum]; 195 196 (*ComputeMBSum)(cur + (i << 4), width, mot_mb); 197 198 totalSAD += mot_mb[0].sad; 199 200 mbnum++; 201 } 202 cur += (width << 4); 203 } 204 205 video->sumMAD = (float)totalSAD / (float)NumPixelMB; 206 207 ResetIntraUpdate(intraArray, totalMB); 208 209 return ; 210 } 211 212 /* 09/20/05 */ 213 if (video->prevBaseVop->padded == 0 && !video->encParams->H263_Enabled) 214 { 215 PaddingEdge(video->prevBaseVop); 216 video->prevBaseVop->padded = 1; 217 } 218 219 /* Random INTRA update */ 220 /* suggest to do it in CodeMB */ 221 /* 2/21/2001 */ 222 //if(video->encParams->RC_Type == CBR_1 || video->encParams->RC_Type == CBR_2) 223 if (video->currLayer == 0 && video->encParams->Refresh) 224 { 225 RasterIntraUpdate(intraArray, Mode, totalMB, video->encParams->Refresh); 226 } 227 228 video->sad_extra_info = NULL; 229 230 #ifdef HTFM 231 /***** HYPOTHESIS TESTING ********/ /* 2/28/01 */ 232 InitHTFM(video, &htfm_stat, newvar, &collect); 233 /*********************************/ 234 #endif 235 236 if ((video->encParams->SceneChange_Det == 1) /*&& video->currLayer==0 */ 237 && ((video->encParams->LayerFrameRate[0] < 5.0) || (video->numVopsInGOP > MIN_GOP))) 238 /* do not try to detect a new scene if low frame rate and too close to previous I-frame */ 239 { 240 incr_i = 2; 241 numLoop = 2; 242 start_i = 1; 243 type_pred = 0; /* for initial candidate selection */ 244 } 245 else 246 { 247 incr_i = 1; 248 numLoop = 1; 249 start_i = 0; 250 type_pred = 2; 251 } 252 253 /* First pass, loop thru half the macroblock */ 254 /* determine scene change */ 255 /* Second pass, for the rest of macroblocks */ 256 numIntra = 0; 257 while (numLoop--) 258 { 259 for (j = 0; j < mbheight; j++) 260 { 261 if (incr_i > 1) 262 start_i = (start_i == 0 ? 1 : 0) ; /* toggle 0 and 1 */ 263 264 offset = width * (j << 4) + (start_i << 4); 265 266 mbnum = j * mbwidth + start_i; 267 268 for (i = start_i; i < mbwidth; i += incr_i) 269 { 270 video->mbnum = mbnum; 271 mot_mb = mot[mbnum]; 272 mode_mb = Mode + mbnum; 273 274 cur = currFrame->yChan + offset; 275 276 277 if (*mode_mb != MODE_INTRA) 278 { 279 #if defined(HTFM) 280 HTFMPrepareCurMB(video, &htfm_stat, cur); 281 #else 282 PrepareCurMB(video, cur); 283 #endif 284 /************************************************************/ 285 /******** full-pel 1MV and 4MVs search **********************/ 286 287 #ifdef _SAD_STAT 288 num_MB++; 289 #endif 290 MBMotionSearch(video, cur, best_cand, i << 4, j << 4, type_pred, 291 FS_en, &hp_guess); 292 293 #ifdef PRINT_MV 294 fp_debug = fopen("c:\\bitstream\\mv1_debug.txt", "a"); 295 fprintf(fp_debug, "#%d (%d,%d,%d) : ", mbnum, mot_mb[0].x, mot_mb[0].y, mot_mb[0].sad); 296 fprintf(fp_debug, "(%d,%d,%d) : (%d,%d,%d) : (%d,%d,%d) : (%d,%d,%d) : ==>\n", 297 mot_mb[1].x, mot_mb[1].y, mot_mb[1].sad, 298 mot_mb[2].x, mot_mb[2].y, mot_mb[2].sad, 299 mot_mb[3].x, mot_mb[3].y, mot_mb[3].sad, 300 mot_mb[4].x, mot_mb[4].y, mot_mb[4].sad); 301 fclose(fp_debug); 302 #endif 303 sad16 = mot_mb[0].sad; 304 #ifdef NO_INTER4V 305 sad8 = sad16; 306 #else 307 sad8 = mot_mb[1].sad + mot_mb[2].sad + mot_mb[3].sad + mot_mb[4].sad; 308 #endif 309 310 /* choose between INTRA or INTER */ 311 (*ChooseMode)(mode_mb, cur, width, ((sad8 < sad16) ? sad8 : sad16)); 312 } 313 else /* INTRA update, use for prediction 3/23/01 */ 314 { 315 mot_mb[0].x = mot_mb[0].y = 0; 316 } 317 318 if (*mode_mb == MODE_INTRA) 319 { 320 numIntra++ ; 321 322 /* compute SAV for rate control and fast DCT, 11/28/00 */ 323 (*ComputeMBSum)(cur, width, mot_mb); 324 325 /* leave mot_mb[0] as it is for fast motion search */ 326 /* set the 4 MVs to zeros */ 327 for (comp = 1; comp <= 4; comp++) 328 { 329 mot_mb[comp].x = 0; 330 mot_mb[comp].y = 0; 331 } 332 #ifdef PRINT_MV 333 fp_debug = fopen("c:\\bitstream\\mv1_debug.txt", "a"); 334 fprintf(fp_debug, "\n"); 335 fclose(fp_debug); 336 #endif 337 } 338 else /* *mode_mb = MODE_INTER;*/ 339 { 340 if (video->encParams->HalfPel_Enabled) 341 { 342 #ifdef _SAD_STAT 343 num_HP_MB++; 344 #endif 345 /* find half-pel resolution motion vector */ 346 FindHalfPelMB(video, cur, mot_mb, best_cand[0], 347 i << 4, j << 4, xh, yh, hp_guess); 348 #ifdef PRINT_MV 349 fp_debug = fopen("c:\\bitstream\\mv1_debug.txt", "a"); 350 fprintf(fp_debug, "(%d,%d), %d\n", mot_mb[0].x, mot_mb[0].y, mot_mb[0].sad); 351 fclose(fp_debug); 352 #endif 353 skip_halfpel_4mv = ((sad16 - mot_mb[0].sad) <= (MB_Nb >> 1) + 1); 354 sad16 = mot_mb[0].sad; 355 356 #ifndef NO_INTER4V 357 if (use_4mv && !skip_halfpel_4mv) 358 { 359 /* Also decide 1MV or 4MV !!!!!!!!*/ 360 sad8 = FindHalfPelBlk(video, cur, mot_mb, sad16, 361 best_cand, mode_mb, i << 4, j << 4, xh, yh, hp_mem4MV); 362 363 #ifdef PRINT_MV 364 fp_debug = fopen("c:\\bitstream\\mv1_debug.txt", "a"); 365 fprintf(fp_debug, " (%d,%d,%d) : (%d,%d,%d) : (%d,%d,%d) : (%d,%d,%d) \n", 366 mot_mb[1].x, mot_mb[1].y, mot_mb[1].sad, 367 mot_mb[2].x, mot_mb[2].y, mot_mb[2].sad, 368 mot_mb[3].x, mot_mb[3].y, mot_mb[3].sad, 369 mot_mb[4].x, mot_mb[4].y, mot_mb[4].sad); 370 fclose(fp_debug); 371 #endif 372 } 373 #endif /* NO_INTER4V */ 374 } 375 else /* HalfPel_Enabled ==0 */ 376 { 377 #ifndef NO_INTER4V 378 //if(sad16 < sad8-PREF_16_VEC) 379 if (sad16 - PREF_16_VEC > sad8) 380 { 381 *mode_mb = MODE_INTER4V; 382 } 383 #endif 384 } 385 #if (ZERO_MV_PREF==2) /* use mot_mb[7].sad as d0 computed in MBMotionSearch*/ 386 /******************************************************/ 387 if (mot_mb[7].sad - PREF_NULL_VEC < sad16 && mot_mb[7].sad - PREF_NULL_VEC < sad8) 388 { 389 mot_mb[0].sad = mot_mb[7].sad - PREF_NULL_VEC; 390 mot_mb[0].x = mot_mb[0].y = 0; 391 *mode_mb = MODE_INTER; 392 } 393 /******************************************************/ 394 #endif 395 if (*mode_mb == MODE_INTER) 396 { 397 if (mot_mb[0].x == 0 && mot_mb[0].y == 0) /* use zero vector */ 398 mot_mb[0].sad += PREF_NULL_VEC; /* add back the bias */ 399 400 mot_mb[1].sad = mot_mb[2].sad = mot_mb[3].sad = mot_mb[4].sad = (mot_mb[0].sad + 2) >> 2; 401 mot_mb[1].x = mot_mb[2].x = mot_mb[3].x = mot_mb[4].x = mot_mb[0].x; 402 mot_mb[1].y = mot_mb[2].y = mot_mb[3].y = mot_mb[4].y = mot_mb[0].y; 403 404 } 405 } 406 407 /* find maximum magnitude */ 408 /* compute average SAD for rate control, 11/28/00 */ 409 if (*mode_mb == MODE_INTER) 410 { 411 #ifdef PRINT_MV 412 fp_debug = fopen("c:\\bitstream\\mv1_debug.txt", "a"); 413 fprintf(fp_debug, "%d MODE_INTER\n", mbnum); 414 fclose(fp_debug); 415 #endif 416 totalSAD += mot_mb[0].sad; 417 if (mot_mb[0].x > max_mag) 418 max_mag = mot_mb[0].x; 419 if (mot_mb[0].y > max_mag) 420 max_mag = mot_mb[0].y; 421 if (mot_mb[0].x < min_mag) 422 min_mag = mot_mb[0].x; 423 if (mot_mb[0].y < min_mag) 424 min_mag = mot_mb[0].y; 425 } 426 else if (*mode_mb == MODE_INTER4V) 427 { 428 #ifdef PRINT_MV 429 fp_debug = fopen("c:\\bitstream\\mv1_debug.txt", "a"); 430 fprintf(fp_debug, "%d MODE_INTER4V\n", mbnum); 431 fclose(fp_debug); 432 #endif 433 totalSAD += sad8; 434 for (comp = 1; comp <= 4; comp++) 435 { 436 if (mot_mb[comp].x > max_mag) 437 max_mag = mot_mb[comp].x; 438 if (mot_mb[comp].y > max_mag) 439 max_mag = mot_mb[comp].y; 440 if (mot_mb[comp].x < min_mag) 441 min_mag = mot_mb[comp].x; 442 if (mot_mb[comp].y < min_mag) 443 min_mag = mot_mb[comp].y; 444 } 445 } 446 else /* MODE_INTRA */ 447 { 448 #ifdef PRINT_MV 449 fp_debug = fopen("c:\\bitstream\\mv1_debug.txt", "a"); 450 fprintf(fp_debug, "%d MODE_INTRA\n", mbnum); 451 fclose(fp_debug); 452 #endif 453 totalSAD += mot_mb[0].sad; 454 } 455 mbnum += incr_i; 456 offset += (incr_i << 4); 457 458 } 459 } 460 461 if (incr_i > 1 && numLoop) /* scene change on and first loop */ 462 { 463 //if(numIntra > ((totalMB>>3)<<1) + (totalMB>>3)) /* 75% of 50%MBs */ 464 if (numIntra > (0.30*(totalMB / 2.0))) /* 15% of 50%MBs */ 465 { 466 /******** scene change detected *******************/ 467 currVop->predictionType = I_VOP; 468 M4VENC_MEMSET(Mode, MODE_INTRA, sizeof(UChar)*totalMB); /* set this for MB level coding*/ 469 currVop->quantizer = video->encParams->InitQuantIvop[video->currLayer]; 470 471 /* compute the SAV for rate control & fast DCT */ 472 totalSAD = 0; 473 offset = 0; 474 mbnum = 0; 475 cur = currFrame->yChan; 476 477 for (j = 0; j < mbheight; j++) 478 { 479 for (i = 0; i < mbwidth; i++) 480 { 481 video->mbnum = mbnum; 482 mot_mb = mot[mbnum]; 483 484 485 (*ComputeMBSum)(cur + (i << 4), width, mot_mb); 486 totalSAD += mot_mb[0].sad; 487 488 mbnum++; 489 } 490 cur += (width << 4); 491 } 492 493 video->sumMAD = (float)totalSAD / (float)NumPixelMB; 494 ResetIntraUpdate(intraArray, totalMB); 495 /* video->numVopsInGOP=0; 3/13/01 move it to vop.c*/ 496 497 return ; 498 } 499 } 500 /******** no scene change, continue motion search **********************/ 501 start_i = 0; 502 type_pred++; /* second pass */ 503 } 504 505 video->sumMAD = (float)totalSAD / (float)NumPixelMB; /* avg SAD */ 506 507 /* find f_code , 10/27/2000 */ 508 f_code_p = 1; 509 while ((max_mag >> (4 + f_code_p)) > 0) 510 f_code_p++; 511 512 f_code_n = 1; 513 min_mag *= -1; 514 while ((min_mag - 1) >> (4 + f_code_n) > 0) 515 f_code_n++; 516 517 currVop->fcodeForward = (f_code_p > f_code_n ? f_code_p : f_code_n); 518 519 #ifdef HTFM 520 /***** HYPOTHESIS TESTING ********/ /* 2/28/01 */ 521 if (collect) 522 { 523 collect = 0; 524 UpdateHTFM(video, newvar, exp_lamda, &htfm_stat); 525 } 526 /*********************************/ 527 #endif 528 529 return ; 530 } 531 532 533 #ifdef HTFM 534 void InitHTFM(VideoEncData *video, HTFM_Stat *htfm_stat, double *newvar, Int *collect) 535 { 536 Int i; 537 Int lx = video->currVop->width; // padding 538 Int lx2 = lx << 1; 539 Int lx3 = lx2 + lx; 540 Int rx = video->currVop->pitch; 541 Int rx2 = rx << 1; 542 Int rx3 = rx2 + rx; 543 544 Int *offset, *offset2; 545 546 /* 4/11/01, collect data every 30 frames, doesn't have to be base layer */ 547 if (((Int)video->numVopsInGOP) % 30 == 1) 548 { 549 550 *collect = 1; 551 552 htfm_stat->countbreak = 0; 553 htfm_stat->abs_dif_mad_avg = 0; 554 555 for (i = 0; i < 16; i++) 556 { 557 newvar[i] = 0.0; 558 } 559 // video->functionPointer->SAD_MB_PADDING = &SAD_MB_PADDING_HTFM_Collect; 560 video->functionPointer->SAD_Macroblock = &SAD_MB_HTFM_Collect; 561 video->functionPointer->SAD_MB_HalfPel[0] = NULL; 562 video->functionPointer->SAD_MB_HalfPel[1] = &SAD_MB_HP_HTFM_Collectxh; 563 video->functionPointer->SAD_MB_HalfPel[2] = &SAD_MB_HP_HTFM_Collectyh; 564 video->functionPointer->SAD_MB_HalfPel[3] = &SAD_MB_HP_HTFM_Collectxhyh; 565 video->sad_extra_info = (void*)(htfm_stat); 566 offset = htfm_stat->offsetArray; 567 offset2 = htfm_stat->offsetRef; 568 } 569 else 570 { 571 // video->functionPointer->SAD_MB_PADDING = &SAD_MB_PADDING_HTFM; 572 video->functionPointer->SAD_Macroblock = &SAD_MB_HTFM; 573 video->functionPointer->SAD_MB_HalfPel[0] = NULL; 574 video->functionPointer->SAD_MB_HalfPel[1] = &SAD_MB_HP_HTFMxh; 575 video->functionPointer->SAD_MB_HalfPel[2] = &SAD_MB_HP_HTFMyh; 576 video->functionPointer->SAD_MB_HalfPel[3] = &SAD_MB_HP_HTFMxhyh; 577 video->sad_extra_info = (void*)(video->nrmlz_th); 578 offset = video->nrmlz_th + 16; 579 offset2 = video->nrmlz_th + 32; 580 } 581 582 offset[0] = 0; 583 offset[1] = lx2 + 2; 584 offset[2] = 2; 585 offset[3] = lx2; 586 offset[4] = lx + 1; 587 offset[5] = lx3 + 3; 588 offset[6] = lx + 3; 589 offset[7] = lx3 + 1; 590 offset[8] = lx; 591 offset[9] = lx3 + 2; 592 offset[10] = lx3 ; 593 offset[11] = lx + 2 ; 594 offset[12] = 1; 595 offset[13] = lx2 + 3; 596 offset[14] = lx2 + 1; 597 offset[15] = 3; 598 599 offset2[0] = 0; 600 offset2[1] = rx2 + 2; 601 offset2[2] = 2; 602 offset2[3] = rx2; 603 offset2[4] = rx + 1; 604 offset2[5] = rx3 + 3; 605 offset2[6] = rx + 3; 606 offset2[7] = rx3 + 1; 607 offset2[8] = rx; 608 offset2[9] = rx3 + 2; 609 offset2[10] = rx3 ; 610 offset2[11] = rx + 2 ; 611 offset2[12] = 1; 612 offset2[13] = rx2 + 3; 613 offset2[14] = rx2 + 1; 614 offset2[15] = 3; 615 616 return ; 617 } 618 619 void UpdateHTFM(VideoEncData *video, double *newvar, double *exp_lamda, HTFM_Stat *htfm_stat) 620 { 621 if (htfm_stat->countbreak == 0) 622 htfm_stat->countbreak = 1; 623 624 newvar[0] = (double)(htfm_stat->abs_dif_mad_avg) / (htfm_stat->countbreak * 16.); 625 626 if (newvar[0] < 0.001) 627 { 628 newvar[0] = 0.001; /* to prevent floating overflow */ 629 } 630 exp_lamda[0] = 1 / (newvar[0] * 1.4142136); 631 exp_lamda[1] = exp_lamda[0] * 1.5825; 632 exp_lamda[2] = exp_lamda[0] * 2.1750; 633 exp_lamda[3] = exp_lamda[0] * 3.5065; 634 exp_lamda[4] = exp_lamda[0] * 3.1436; 635 exp_lamda[5] = exp_lamda[0] * 3.5315; 636 exp_lamda[6] = exp_lamda[0] * 3.7449; 637 exp_lamda[7] = exp_lamda[0] * 4.5854; 638 exp_lamda[8] = exp_lamda[0] * 4.6191; 639 exp_lamda[9] = exp_lamda[0] * 5.4041; 640 exp_lamda[10] = exp_lamda[0] * 6.5974; 641 exp_lamda[11] = exp_lamda[0] * 10.5341; 642 exp_lamda[12] = exp_lamda[0] * 10.0719; 643 exp_lamda[13] = exp_lamda[0] * 12.0516; 644 exp_lamda[14] = exp_lamda[0] * 15.4552; 645 646 CalcThreshold(HTFM_Pf, exp_lamda, video->nrmlz_th); 647 return ; 648 } 649 650 651 void CalcThreshold(double pf, double exp_lamda[], Int nrmlz_th[]) 652 { 653 Int i; 654 double temp[15]; 655 // printf("\nLamda: "); 656 657 /* parametric PREMODELling */ 658 for (i = 0; i < 15; i++) 659 { 660 // printf("%g ",exp_lamda[i]); 661 if (pf < 0.5) 662 temp[i] = 1 / exp_lamda[i] * M4VENC_LOG(2 * pf); 663 else 664 temp[i] = -1 / exp_lamda[i] * M4VENC_LOG(2 * (1 - pf)); 665 } 666 667 nrmlz_th[15] = 0; 668 for (i = 0; i < 15; i++) /* scale upto no.pixels */ 669 nrmlz_th[i] = (Int)(temp[i] * ((i + 1) << 4) + 0.5); 670 671 return ; 672 } 673 674 void HTFMPrepareCurMB(VideoEncData *video, HTFM_Stat *htfm_stat, UChar *cur) 675 { 676 void* tmp = (void*)(video->currYMB); 677 ULong *htfmMB = (ULong*)tmp; 678 UChar *ptr, byte; 679 Int *offset; 680 Int i; 681 ULong word; 682 Int width = video->currVop->width; 683 684 if (((Int)video->numVopsInGOP) % 30 == 1) 685 { 686 offset = htfm_stat->offsetArray; 687 } 688 else 689 { 690 offset = video->nrmlz_th + 16; 691 } 692 693 for (i = 0; i < 16; i++) 694 { 695 ptr = cur + offset[i]; 696 word = ptr[0]; 697 byte = ptr[4]; 698 word |= (byte << 8); 699 byte = ptr[8]; 700 word |= (byte << 16); 701 byte = ptr[12]; 702 word |= (byte << 24); 703 *htfmMB++ = word; 704 705 word = *(ptr += (width << 2)); 706 byte = ptr[4]; 707 word |= (byte << 8); 708 byte = ptr[8]; 709 word |= (byte << 16); 710 byte = ptr[12]; 711 word |= (byte << 24); 712 *htfmMB++ = word; 713 714 word = *(ptr += (width << 2)); 715 byte = ptr[4]; 716 word |= (byte << 8); 717 byte = ptr[8]; 718 word |= (byte << 16); 719 byte = ptr[12]; 720 word |= (byte << 24); 721 *htfmMB++ = word; 722 723 word = *(ptr += (width << 2)); 724 byte = ptr[4]; 725 word |= (byte << 8); 726 byte = ptr[8]; 727 word |= (byte << 16); 728 byte = ptr[12]; 729 word |= (byte << 24); 730 *htfmMB++ = word; 731 } 732 733 return ; 734 } 735 736 737 #endif 738 739 void PrepareCurMB(VideoEncData *video, UChar *cur) 740 { 741 void* tmp = (void*)(video->currYMB); 742 ULong *currYMB = (ULong*)tmp; 743 Int i; 744 Int width = video->currVop->width; 745 746 cur -= width; 747 748 for (i = 0; i < 16; i++) 749 { 750 *currYMB++ = *((ULong*)(cur += width)); 751 *currYMB++ = *((ULong*)(cur + 4)); 752 *currYMB++ = *((ULong*)(cur + 8)); 753 *currYMB++ = *((ULong*)(cur + 12)); 754 } 755 756 return ; 757 } 758 759 760 /*================================================================== 761 Function: MBMotionSearch 762 Date: 09/06/2000 763 Purpose: Perform motion estimation for a macroblock. 764 Find 1MV and 4MVs in half-pels resolutions. 765 Using ST1 algorithm provided by Chalidabhongse and Kuo 766 CSVT March'98. 767 768 ==================================================================*/ 769 770 void MBMotionSearch(VideoEncData *video, UChar *cur, UChar *best_cand[], 771 Int i0, Int j0, Int type_pred, Int FS_en, Int *hp_guess) 772 { 773 Vol *currVol = video->vol[video->currLayer]; 774 UChar *ref, *cand, *ncand = NULL, *cur8; 775 void *extra_info = video->sad_extra_info; 776 Int mbnum = video->mbnum; 777 Int width = video->currVop->width; /* 6/12/01, must be multiple of 16 */ 778 Int height = video->currVop->height; 779 MOT **mot = video->mot; 780 UChar use_4mv = video->encParams->MV8x8_Enabled; 781 UChar h263_mode = video->encParams->H263_Enabled; 782 Int(*SAD_Macroblock)(UChar*, UChar*, Int, void*) = video->functionPointer->SAD_Macroblock; 783 Int(*SAD_Block)(UChar*, UChar*, Int, Int, void*) = video->functionPointer->SAD_Block; 784 VideoEncParams *encParams = video->encParams; 785 Int range = encParams->SearchRange; 786 787 Int lx = video->currVop->pitch; /* padding */ 788 Int comp; 789 Int i, j, imin, jmin, ilow, ihigh, jlow, jhigh, iorg, jorg; 790 Int d, dmin, dn[9]; 791 #if (ZERO_MV_PREF==1) /* compute (0,0) MV at the end */ 792 Int d0; 793 #endif 794 Int k; 795 Int mvx[5], mvy[5], imin0, jmin0; 796 Int num_can, center_again; 797 Int last_loc, new_loc = 0; 798 Int step, max_step = range >> 1; 799 Int next; 800 801 ref = video->forwardRefVop->yChan; /* origin of actual frame */ 802 803 cur = video->currYMB; /* use smaller memory space for current MB */ 804 805 /* find limit of the search (adjusting search range)*/ 806 807 if (!h263_mode) 808 { 809 ilow = i0 - range; 810 if (ilow < -15) 811 ilow = -15; 812 ihigh = i0 + range - 1; 813 if (ihigh > width - 1) 814 ihigh = width - 1; 815 jlow = j0 - range; 816 if (jlow < -15) 817 jlow = -15; 818 jhigh = j0 + range - 1; 819 if (jhigh > height - 1) 820 jhigh = height - 1; 821 } 822 else 823 { 824 ilow = i0 - range; 825 if (ilow < 0) 826 ilow = 0; 827 ihigh = i0 + range - 1; 828 if (ihigh > width - 16) 829 ihigh = width - 16; 830 jlow = j0 - range; 831 if (jlow < 0) 832 jlow = 0; 833 jhigh = j0 + range - 1; 834 if (jhigh > height - 16) 835 jhigh = height - 16; 836 } 837 838 imin = i0; 839 jmin = j0; /* needed for fullsearch */ 840 ncand = ref + imin + jmin * lx; 841 842 /* for first row of MB, fullsearch can be used */ 843 if (FS_en) 844 { 845 *hp_guess = 0; /* no guess for fast half-pel */ 846 847 dmin = fullsearch(video, currVol, ref, cur, &imin, &jmin, ilow, ihigh, jlow, jhigh); 848 849 ncand = ref + imin + jmin * lx; 850 851 mot[mbnum][0].sad = dmin; 852 mot[mbnum][0].x = (imin - i0) << 1; 853 mot[mbnum][0].y = (jmin - j0) << 1; 854 imin0 = imin << 1; /* 16x16 MV in half-pel resolution */ 855 jmin0 = jmin << 1; 856 best_cand[0] = ncand; 857 } 858 else 859 { /* 4/7/01, modified this testing for fullsearch the top row to only upto (0,3) MB */ 860 /* upto 30% complexity saving with the same complexity */ 861 if (video->forwardRefVop->predictionType == I_VOP && j0 == 0 && i0 <= 64 && type_pred != 1) 862 { 863 *hp_guess = 0; /* no guess for fast half-pel */ 864 dmin = fullsearch(video, currVol, ref, cur, &imin, &jmin, ilow, ihigh, jlow, jhigh); 865 ncand = ref + imin + jmin * lx; 866 } 867 else 868 { 869 /************** initialize candidate **************************/ 870 /* find initial motion vector */ 871 CandidateSelection(mvx, mvy, &num_can, i0 >> 4, j0 >> 4, video, type_pred); 872 873 dmin = 65535; 874 875 /* check if all are equal */ 876 if (num_can == ALL_CAND_EQUAL) 877 { 878 i = i0 + mvx[0]; 879 j = j0 + mvy[0]; 880 881 if (i >= ilow && i <= ihigh && j >= jlow && j <= jhigh) 882 { 883 cand = ref + i + j * lx; 884 885 d = (*SAD_Macroblock)(cand, cur, (dmin << 16) | lx, extra_info); 886 887 if (d < dmin) 888 { 889 dmin = d; 890 imin = i; 891 jmin = j; 892 ncand = cand; 893 } 894 } 895 } 896 else 897 { 898 /************** evaluate unique candidates **********************/ 899 for (k = 0; k < num_can; k++) 900 { 901 i = i0 + mvx[k]; 902 j = j0 + mvy[k]; 903 904 if (i >= ilow && i <= ihigh && j >= jlow && j <= jhigh) 905 { 906 cand = ref + i + j * lx; 907 d = (*SAD_Macroblock)(cand, cur, (dmin << 16) | lx, extra_info); 908 909 if (d < dmin) 910 { 911 dmin = d; 912 imin = i; 913 jmin = j; 914 ncand = cand; 915 } 916 else if ((d == dmin) && PV_ABS(mvx[k]) + PV_ABS(mvy[k]) < PV_ABS(i0 - imin) + PV_ABS(j0 - jmin)) 917 { 918 dmin = d; 919 imin = i; 920 jmin = j; 921 ncand = cand; 922 } 923 } 924 } 925 } 926 if (num_can == 0 || dmin == 65535) /* no candidate selected */ 927 { 928 ncand = ref + i0 + j0 * lx; /* use (0,0) MV as initial value */ 929 mot[mbnum][7].sad = dmin = (*SAD_Macroblock)(ncand, cur, (65535 << 16) | lx, extra_info); 930 #if (ZERO_MV_PREF==1) /* compute (0,0) MV at the end */ 931 d0 = dmin; 932 #endif 933 imin = i0; 934 jmin = j0; 935 } 936 937 #if (ZERO_MV_PREF==0) /* COMPUTE ZERO VECTOR FIRST !!!!!*/ 938 dmin -= PREF_NULL_VEC; 939 #endif 940 941 /******************* local refinement ***************************/ 942 center_again = 0; 943 last_loc = new_loc = 0; 944 // ncand = ref + jmin*lx + imin; /* center of the search */ 945 step = 0; 946 dn[0] = dmin; 947 while (!center_again && step <= max_step) 948 { 949 950 MoveNeighborSAD(dn, last_loc); 951 952 center_again = 1; 953 i = imin; 954 j = jmin - 1; 955 cand = ref + i + j * lx; 956 957 /* starting from [0,-1] */ 958 /* spiral check one step at a time*/ 959 for (k = 2; k <= 8; k += 2) 960 { 961 if (!tab_exclude[last_loc][k]) /* exclude last step computation */ 962 { /* not already computed */ 963 if (i >= ilow && i <= ihigh && j >= jlow && j <= jhigh) 964 { 965 d = (*SAD_Macroblock)(cand, cur, (dmin << 16) | lx, extra_info); 966 dn[k] = d; /* keep it for half pel use */ 967 968 if (d < dmin) 969 { 970 ncand = cand; 971 dmin = d; 972 imin = i; 973 jmin = j; 974 center_again = 0; 975 new_loc = k; 976 } 977 else if ((d == dmin) && PV_ABS(i0 - i) + PV_ABS(j0 - j) < PV_ABS(i0 - imin) + PV_ABS(j0 - jmin)) 978 { 979 ncand = cand; 980 imin = i; 981 jmin = j; 982 center_again = 0; 983 new_loc = k; 984 } 985 } 986 } 987 if (k == 8) /* end side search*/ 988 { 989 if (!center_again) 990 { 991 k = -1; /* start diagonal search */ 992 cand -= lx; 993 j--; 994 } 995 } 996 else 997 { 998 next = refine_next[k][0]; 999 i += next; 1000 cand += next; 1001 next = refine_next[k][1]; 1002 j += next; 1003 cand += lx * next; 1004 } 1005 } 1006 last_loc = new_loc; 1007 step ++; 1008 } 1009 if (!center_again) 1010 MoveNeighborSAD(dn, last_loc); 1011 1012 *hp_guess = FindMin(dn); 1013 1014 } 1015 1016 #if (ZERO_MV_PREF==1) /* compute (0,0) MV at the end */ 1017 if (d0 - PREF_NULL_VEC < dmin) 1018 { 1019 ncand = ref + i0 + j0 * lx; 1020 dmin = d0; 1021 imin = i0; 1022 jmin = j0; 1023 } 1024 #endif 1025 mot[mbnum][0].sad = dmin; 1026 mot[mbnum][0].x = (imin - i0) << 1; 1027 mot[mbnum][0].y = (jmin - j0) << 1; 1028 imin0 = imin << 1; /* 16x16 MV in half-pel resolution */ 1029 jmin0 = jmin << 1; 1030 best_cand[0] = ncand; 1031 } 1032 /* imin and jmin is the best 1 MV */ 1033 #ifndef NO_INTER4V 1034 /******************* Find 4 motion vectors ****************************/ 1035 if (use_4mv && !h263_mode) 1036 { 1037 #ifdef _SAD_STAT 1038 num_Blk += 4; 1039 #endif 1040 /* starting from the best 1MV */ 1041 //offset = imin + jmin*lx; 1042 iorg = i0; 1043 jorg = j0; 1044 1045 for (comp = 0; comp < 4; comp++) 1046 { 1047 i0 = iorg + ((comp & 1) << 3); 1048 j0 = jorg + ((comp & 2) << 2); 1049 1050 imin = (imin0 >> 1) + ((comp & 1) << 3); /* starting point from 16x16 MV */ 1051 jmin = (jmin0 >> 1) + ((comp & 2) << 2); 1052 ncand = ref + imin + jmin * lx; 1053 1054 cur8 = cur + ((comp & 1) << 3) + (((comp & 2) << 2) << 4) ; /* 11/30/05, smaller cache */ 1055 1056 /* find limit of the search (adjusting search range)*/ 1057 ilow = i0 - range; 1058 ihigh = i0 + range - 1 ;/* 4/9/01 */ 1059 if (ilow < -15) 1060 ilow = -15; 1061 if (ihigh > width - 1) 1062 ihigh = width - 1; 1063 jlow = j0 - range; 1064 jhigh = j0 + range - 1 ;/* 4/9/01 */ 1065 if (jlow < -15) 1066 jlow = -15; 1067 if (jhigh > height - 1) 1068 jhigh = height - 1; 1069 1070 SAD_Block = video->functionPointer->SAD_Block; 1071 1072 if (FS_en) /* fullsearch enable, center around 16x16 MV */ 1073 { 1074 dmin = fullsearchBlk(video, currVol, ncand, cur8, &imin, &jmin, ilow, ihigh, jlow, jhigh, range); 1075 ncand = ref + imin + jmin * lx; 1076 1077 mot[mbnum][comp+1].sad = dmin; 1078 mot[mbnum][comp+1].x = (imin - i0) << 1; 1079 mot[mbnum][comp+1].y = (jmin - j0) << 1; 1080 best_cand[comp+1] = ncand; 1081 } 1082 else /* no fullsearch, do local search */ 1083 { 1084 /* starting point from 16x16 */ 1085 dmin = (*SAD_Block)(ncand, cur8, 65536, lx, extra_info); 1086 1087 /******************* local refinement ***************************/ 1088 center_again = 0; 1089 last_loc = 0; 1090 1091 while (!center_again) 1092 { 1093 center_again = 1; 1094 i = imin; 1095 j = jmin - 1; 1096 cand = ref + i + j * lx; 1097 1098 /* starting from [0,-1] */ 1099 /* spiral check one step at a time*/ 1100 for (k = 2; k <= 8; k += 2) 1101 { 1102 if (!tab_exclude[last_loc][k]) /* exclude last step computation */ 1103 { /* not already computed */ 1104 if (i >= ilow && i <= ihigh && j >= jlow && j <= jhigh) 1105 { 1106 d = (*SAD_Block)(cand, cur8, dmin, lx, extra_info); 1107 1108 if (d < dmin) 1109 { 1110 ncand = cand; 1111 dmin = d; 1112 imin = i; 1113 jmin = j; 1114 center_again = 0; 1115 new_loc = k; 1116 } 1117 else if ((d == dmin) && 1118 PV_ABS(i0 - i) + PV_ABS(j0 - j) < PV_ABS(i0 - imin) + PV_ABS(j0 - jmin)) 1119 { 1120 ncand = cand; 1121 imin = i; 1122 jmin = j; 1123 center_again = 0; 1124 new_loc = k; 1125 } 1126 } 1127 } 1128 if (k == 8) /* end side search*/ 1129 { 1130 if (!center_again) 1131 { 1132 k = -1; /* start diagonal search */ 1133 if (j <= height - 1 && j > 0) cand -= lx; 1134 j--; 1135 } 1136 } 1137 else 1138 { 1139 next = refine_next[k][0]; 1140 cand += next; 1141 i += next; 1142 next = refine_next[k][1]; 1143 cand += lx * next; 1144 j += next; 1145 } 1146 } 1147 last_loc = new_loc; 1148 } 1149 mot[mbnum][comp+1].sad = dmin; 1150 mot[mbnum][comp+1].x = (imin - i0) << 1; 1151 mot[mbnum][comp+1].y = (jmin - j0) << 1; 1152 best_cand[comp+1] = ncand; 1153 } 1154 /********************************************/ 1155 } 1156 } 1157 else 1158 #endif /* NO_INTER4V */ 1159 { 1160 mot[mbnum][1].sad = mot[mbnum][2].sad = mot[mbnum][3].sad = mot[mbnum][4].sad = (dmin + 2) >> 2; 1161 mot[mbnum][1].x = mot[mbnum][2].x = mot[mbnum][3].x = mot[mbnum][4].x = mot[mbnum][0].x; 1162 mot[mbnum][1].y = mot[mbnum][2].y = mot[mbnum][3].y = mot[mbnum][4].y = mot[mbnum][0].y; 1163 best_cand[1] = best_cand[2] = best_cand[3] = best_cand[4] = ncand; 1164 1165 } 1166 return ; 1167 } 1168 1169 1170 /*=============================================================================== 1171 Function: fullsearch 1172 Date: 09/16/2000 1173 Purpose: Perform full-search motion estimation over the range of search 1174 region in a spiral-outward manner. 1175 Input/Output: VideoEncData, current Vol, previou Vop, pointer to the left corner of 1176 current VOP, current coord (also output), boundaries. 1177 ===============================================================================*/ 1178 1179 Int fullsearch(VideoEncData *video, Vol *currVol, UChar *prev, UChar *cur, 1180 Int *imin, Int *jmin, Int ilow, Int ihigh, Int jlow, Int jhigh) 1181 { 1182 Int range = video->encParams->SearchRange; 1183 UChar *cand; 1184 Int i, j, k, l; 1185 Int d, dmin; 1186 Int i0 = *imin; /* current position */ 1187 Int j0 = *jmin; 1188 Int(*SAD_Macroblock)(UChar*, UChar*, Int, void*) = video->functionPointer->SAD_Macroblock; 1189 void *extra_info = video->sad_extra_info; 1190 // UChar h263_mode = video->encParams->H263_Enabled; 1191 Int lx = video->currVop->pitch; /* with padding */ 1192 1193 Int offset = i0 + j0 * lx; 1194 1195 OSCL_UNUSED_ARG(currVol); 1196 1197 cand = prev + offset; 1198 1199 dmin = (*SAD_Macroblock)(cand, cur, (65535 << 16) | lx, (void*)extra_info) - PREF_NULL_VEC; 1200 1201 /* perform spiral search */ 1202 for (k = 1; k <= range; k++) 1203 { 1204 1205 i = i0 - k; 1206 j = j0 - k; 1207 1208 cand = prev + i + j * lx; 1209 1210 for (l = 0; l < 8*k; l++) 1211 { 1212 /* no need for boundary checking again */ 1213 if (i >= ilow && i <= ihigh && j >= jlow && j <= jhigh) 1214 { 1215 d = (*SAD_Macroblock)(cand, cur, (dmin << 16) | lx, (void*)extra_info); 1216 1217 if (d < dmin) 1218 { 1219 dmin = d; 1220 *imin = i; 1221 *jmin = j; 1222 } 1223 else if ((d == dmin) && PV_ABS(i0 - i) + PV_ABS(j0 - j) < PV_ABS(i0 - *imin) + PV_ABS(j0 - *jmin)) 1224 { 1225 dmin = d; 1226 *imin = i; 1227 *jmin = j; 1228 } 1229 } 1230 1231 if (l < (k << 1)) 1232 { 1233 i++; 1234 cand++; 1235 } 1236 else if (l < (k << 2)) 1237 { 1238 j++; 1239 cand += lx; 1240 } 1241 else if (l < ((k << 2) + (k << 1))) 1242 { 1243 i--; 1244 cand--; 1245 } 1246 else 1247 { 1248 j--; 1249 cand -= lx; 1250 } 1251 } 1252 } 1253 1254 return dmin; 1255 } 1256 1257 #ifndef NO_INTER4V 1258 /*=============================================================================== 1259 Function: fullsearchBlk 1260 Date: 01/9/2001 1261 Purpose: Perform full-search motion estimation of an 8x8 block over the range 1262 of search region in a spiral-outward manner centered at the 16x16 MV. 1263 Input/Output: VideoEncData, MB coordinate, pointer to the initial MV on the 1264 reference, pointer to coor of current block, search range. 1265 ===============================================================================*/ 1266 Int fullsearchBlk(VideoEncData *video, Vol *currVol, UChar *cent, UChar *cur, 1267 Int *imin, Int *jmin, Int ilow, Int ihigh, Int jlow, Int jhigh, Int range) 1268 { 1269 UChar *cand, *ref; 1270 Int i, j, k, l, istart, jstart; 1271 Int d, dmin; 1272 Int lx = video->currVop->pitch; /* with padding */ 1273 Int(*SAD_Block)(UChar*, UChar*, Int, Int, void*) = video->functionPointer->SAD_Block; 1274 void *extra_info = video->sad_extra_info; 1275 1276 OSCL_UNUSED_ARG(currVol); 1277 1278 /* starting point centered at 16x16 MV */ 1279 ref = cent; 1280 istart = *imin; 1281 jstart = *jmin; 1282 1283 dmin = (*SAD_Block)(ref, cur, 65536, lx, (void*)extra_info); 1284 1285 cand = ref; 1286 /* perform spiral search */ 1287 for (k = 1; k <= range; k++) 1288 { 1289 1290 i = istart - k; 1291 j = jstart - k; 1292 cand -= (lx + 1); /* candidate region */ 1293 1294 for (l = 0; l < 8*k; l++) 1295 { 1296 /* no need for boundary checking again */ 1297 if (i >= ilow && i <= ihigh && j >= jlow && j <= jhigh) 1298 { 1299 d = (*SAD_Block)(cand, cur, dmin, lx, (void*)extra_info); 1300 1301 if (d < dmin) 1302 { 1303 dmin = d; 1304 *imin = i; 1305 *jmin = j; 1306 } 1307 else if ((d == dmin) && 1308 PV_ABS(istart - i) + PV_ABS(jstart - j) < PV_ABS(istart - *imin) + PV_ABS(jstart - *jmin)) 1309 { 1310 dmin = d; 1311 *imin = i; 1312 *jmin = j; 1313 } 1314 } 1315 1316 if (l < (k << 1)) 1317 { 1318 i++; 1319 cand++; 1320 } 1321 else if (l < (k << 2)) 1322 { 1323 j++; 1324 cand += lx; 1325 } 1326 else if (l < ((k << 2) + (k << 1))) 1327 { 1328 i--; 1329 cand--; 1330 } 1331 else 1332 { 1333 j--; 1334 cand -= lx; 1335 } 1336 } 1337 } 1338 1339 return dmin; 1340 } 1341 #endif /* NO_INTER4V */ 1342 1343 /*=============================================================================== 1344 Function: CandidateSelection 1345 Date: 09/16/2000 1346 Purpose: Fill up the list of candidate using spatio-temporal correlation 1347 among neighboring blocks. 1348 Input/Output: type_pred = 0: first pass, 1: second pass, or no SCD 1349 Modified: 09/23/01, get rid of redundant candidates before passing back. 1350 ===============================================================================*/ 1351 1352 void CandidateSelection(Int *mvx, Int *mvy, Int *num_can, Int imb, Int jmb, 1353 VideoEncData *video, Int type_pred) 1354 { 1355 MOT **mot = video->mot; 1356 MOT *pmot; 1357 Int mbnum = video->mbnum; 1358 Vol *currVol = video->vol[video->currLayer]; 1359 Int mbwidth = currVol->nMBPerRow; 1360 Int mbheight = currVol->nMBPerCol; 1361 Int i, j, same, num1; 1362 1363 *num_can = 0; 1364 1365 if (video->forwardRefVop->predictionType == P_VOP) 1366 { 1367 /* Spatio-Temporal Candidate (five candidates) */ 1368 if (type_pred == 0) /* first pass */ 1369 { 1370 pmot = &mot[mbnum][0]; /* same coordinate previous frame */ 1371 mvx[(*num_can)] = (pmot->x) >> 1; 1372 mvy[(*num_can)++] = (pmot->y) >> 1; 1373 if (imb >= (mbwidth >> 1) && imb > 0) /*left neighbor previous frame */ 1374 { 1375 pmot = &mot[mbnum-1][0]; 1376 mvx[(*num_can)] = (pmot->x) >> 1; 1377 mvy[(*num_can)++] = (pmot->y) >> 1; 1378 } 1379 else if (imb + 1 < mbwidth) /*right neighbor previous frame */ 1380 { 1381 pmot = &mot[mbnum+1][0]; 1382 mvx[(*num_can)] = (pmot->x) >> 1; 1383 mvy[(*num_can)++] = (pmot->y) >> 1; 1384 } 1385 1386 if (jmb < mbheight - 1) /*bottom neighbor previous frame */ 1387 { 1388 pmot = &mot[mbnum+mbwidth][0]; 1389 mvx[(*num_can)] = (pmot->x) >> 1; 1390 mvy[(*num_can)++] = (pmot->y) >> 1; 1391 } 1392 else if (jmb > 0) /*upper neighbor previous frame */ 1393 { 1394 pmot = &mot[mbnum-mbwidth][0]; 1395 mvx[(*num_can)] = (pmot->x) >> 1; 1396 mvy[(*num_can)++] = (pmot->y) >> 1; 1397 } 1398 1399 if (imb > 0 && jmb > 0) /* upper-left neighbor current frame*/ 1400 { 1401 pmot = &mot[mbnum-mbwidth-1][0]; 1402 mvx[(*num_can)] = (pmot->x) >> 1; 1403 mvy[(*num_can)++] = (pmot->y) >> 1; 1404 } 1405 if (jmb > 0 && imb < mbheight - 1) /* upper right neighbor current frame*/ 1406 { 1407 pmot = &mot[mbnum-mbwidth+1][0]; 1408 mvx[(*num_can)] = (pmot->x) >> 1; 1409 mvy[(*num_can)++] = (pmot->y) >> 1; 1410 } 1411 } 1412 else /* second pass */ 1413 /* original ST1 algorithm */ 1414 { 1415 pmot = &mot[mbnum][0]; /* same coordinate previous frame */ 1416 mvx[(*num_can)] = (pmot->x) >> 1; 1417 mvy[(*num_can)++] = (pmot->y) >> 1; 1418 1419 if (imb > 0) /*left neighbor current frame */ 1420 { 1421 pmot = &mot[mbnum-1][0]; 1422 mvx[(*num_can)] = (pmot->x) >> 1; 1423 mvy[(*num_can)++] = (pmot->y) >> 1; 1424 } 1425 if (jmb > 0) /*upper neighbor current frame */ 1426 { 1427 pmot = &mot[mbnum-mbwidth][0]; 1428 mvx[(*num_can)] = (pmot->x) >> 1; 1429 mvy[(*num_can)++] = (pmot->y) >> 1; 1430 } 1431 if (imb < mbwidth - 1) /*right neighbor previous frame */ 1432 { 1433 pmot = &mot[mbnum+1][0]; 1434 mvx[(*num_can)] = (pmot->x) >> 1; 1435 mvy[(*num_can)++] = (pmot->y) >> 1; 1436 } 1437 if (jmb < mbheight - 1) /*bottom neighbor previous frame */ 1438 { 1439 pmot = &mot[mbnum+mbwidth][0]; 1440 mvx[(*num_can)] = (pmot->x) >> 1; 1441 mvy[(*num_can)++] = (pmot->y) >> 1; 1442 } 1443 } 1444 } 1445 else /* only Spatial Candidate (four candidates)*/ 1446 { 1447 if (type_pred == 0) /*first pass*/ 1448 { 1449 if (imb > 1) /* neighbor two blocks away to the left */ 1450 { 1451 pmot = &mot[mbnum-2][0]; 1452 mvx[(*num_can)] = (pmot->x) >> 1; 1453 mvy[(*num_can)++] = (pmot->y) >> 1; 1454 } 1455 if (imb > 0 && jmb > 0) /* upper-left neighbor */ 1456 { 1457 pmot = &mot[mbnum-mbwidth-1][0]; 1458 mvx[(*num_can)] = (pmot->x) >> 1; 1459 mvy[(*num_can)++] = (pmot->y) >> 1; 1460 } 1461 if (jmb > 0 && imb < mbheight - 1) /* upper right neighbor */ 1462 { 1463 pmot = &mot[mbnum-mbwidth+1][0]; 1464 mvx[(*num_can)] = (pmot->x) >> 1; 1465 mvy[(*num_can)++] = (pmot->y) >> 1; 1466 } 1467 } 1468 //#ifdef SCENE_CHANGE_DETECTION 1469 /* second pass (ST2 algorithm)*/ 1470 else if (type_pred == 1) /* 4/7/01 */ 1471 { 1472 if (imb > 0) /*left neighbor current frame */ 1473 { 1474 pmot = &mot[mbnum-1][0]; 1475 mvx[(*num_can)] = (pmot->x) >> 1; 1476 mvy[(*num_can)++] = (pmot->y) >> 1; 1477 } 1478 if (jmb > 0) /*upper neighbor current frame */ 1479 { 1480 pmot = &mot[mbnum-mbwidth][0]; 1481 mvx[(*num_can)] = (pmot->x) >> 1; 1482 mvy[(*num_can)++] = (pmot->y) >> 1; 1483 } 1484 if (imb < mbwidth - 1) /*right neighbor current frame */ 1485 { 1486 pmot = &mot[mbnum+1][0]; 1487 mvx[(*num_can)] = (pmot->x) >> 1; 1488 mvy[(*num_can)++] = (pmot->y) >> 1; 1489 } 1490 if (jmb < mbheight - 1) /*bottom neighbor current frame */ 1491 { 1492 pmot = &mot[mbnum+mbwidth][0]; 1493 mvx[(*num_can)] = (pmot->x) >> 1; 1494 mvy[(*num_can)++] = (pmot->y) >> 1; 1495 } 1496 } 1497 //#else 1498 else /* original ST1 algorithm */ 1499 { 1500 if (imb > 0) /*left neighbor current frame */ 1501 { 1502 pmot = &mot[mbnum-1][0]; 1503 mvx[(*num_can)] = (pmot->x) >> 1; 1504 mvy[(*num_can)++] = (pmot->y) >> 1; 1505 1506 if (jmb > 0) /*upper-left neighbor current frame */ 1507 { 1508 pmot = &mot[mbnum-mbwidth-1][0]; 1509 mvx[(*num_can)] = (pmot->x) >> 1; 1510 mvy[(*num_can)++] = (pmot->y) >> 1; 1511 } 1512 1513 } 1514 if (jmb > 0) /*upper neighbor current frame */ 1515 { 1516 pmot = &mot[mbnum-mbwidth][0]; 1517 mvx[(*num_can)] = (pmot->x) >> 1; 1518 mvy[(*num_can)++] = (pmot->y) >> 1; 1519 1520 if (imb < mbheight - 1) /*upper-right neighbor current frame */ 1521 { 1522 pmot = &mot[mbnum-mbwidth+1][0]; 1523 mvx[(*num_can)] = (pmot->x) >> 1; 1524 mvy[(*num_can)++] = (pmot->y) >> 1; 1525 } 1526 } 1527 } 1528 //#endif 1529 } 1530 1531 /* 3/23/01, remove redundant candidate (possible k-mean) */ 1532 num1 = *num_can; 1533 *num_can = 1; 1534 for (i = 1; i < num1; i++) 1535 { 1536 same = 0; 1537 j = 0; 1538 while (!same && j < *num_can) 1539 { 1540 #if (CANDIDATE_DISTANCE==0) 1541 if (mvx[i] == mvx[j] && mvy[i] == mvy[j]) 1542 #else 1543 // modified k-mean, 3/24/01, shouldn't be greater than 3 1544 if (PV_ABS(mvx[i] - mvx[j]) + PV_ABS(mvy[i] - mvy[j]) < CANDIDATE_DISTANCE) 1545 #endif 1546 same = 1; 1547 j++; 1548 } 1549 if (!same) 1550 { 1551 mvx[*num_can] = mvx[i]; 1552 mvy[*num_can] = mvy[i]; 1553 (*num_can)++; 1554 } 1555 } 1556 1557 #ifdef _SAD_STAT 1558 num_cand += (*num_can); 1559 #endif 1560 1561 if (num1 == 5 && *num_can == 1) 1562 *num_can = ALL_CAND_EQUAL; /* all are equal */ 1563 1564 return ; 1565 } 1566 1567 /*=========================================================================== 1568 Function: RasterIntraUpdate 1569 Date: 2/26/01 1570 Purpose: To raster-scan assign INTRA-update . 1571 N macroblocks are updated (also was programmable). 1572 ===========================================================================*/ 1573 void RasterIntraUpdate(UChar *intraArray, UChar *Mode, Int totalMB, Int numRefresh) 1574 { 1575 Int indx, i; 1576 1577 /* find the last refresh MB */ 1578 indx = 0; 1579 while (intraArray[indx] == 1 && indx < totalMB) 1580 indx++; 1581 1582 /* add more */ 1583 for (i = 0; i < numRefresh && indx < totalMB; i++) 1584 { 1585 Mode[indx] = MODE_INTRA; 1586 intraArray[indx++] = 1; 1587 } 1588 1589 /* if read the end of frame, reset and loop around */ 1590 if (indx >= totalMB - 1) 1591 { 1592 ResetIntraUpdate(intraArray, totalMB); 1593 indx = 0; 1594 while (i < numRefresh && indx < totalMB) 1595 { 1596 intraArray[indx] = 1; 1597 Mode[indx++] = MODE_INTRA; 1598 i++; 1599 } 1600 } 1601 1602 return ; 1603 } 1604 1605 /*=========================================================================== 1606 Function: ResetIntraUpdate 1607 Date: 11/28/00 1608 Purpose: Reset already intra updated flags to all zero 1609 ===========================================================================*/ 1610 1611 void ResetIntraUpdate(UChar *intraArray, Int totalMB) 1612 { 1613 M4VENC_MEMSET(intraArray, 0, sizeof(UChar)*totalMB); 1614 return ; 1615 } 1616 1617 /*=========================================================================== 1618 Function: ResetIntraUpdateRegion 1619 Date: 12/1/00 1620 Purpose: Reset already intra updated flags in one region to all zero 1621 ===========================================================================*/ 1622 void ResetIntraUpdateRegion(UChar *intraArray, Int start_i, Int rwidth, 1623 Int start_j, Int rheight, Int mbwidth, Int mbheight) 1624 { 1625 Int indx, j; 1626 1627 if (start_i + rwidth >= mbwidth) 1628 rwidth = mbwidth - start_i; 1629 if (start_j + rheight >= mbheight) 1630 rheight = mbheight - start_j; 1631 1632 for (j = start_j; j < start_j + rheight; j++) 1633 { 1634 indx = j * mbwidth; 1635 M4VENC_MEMSET(intraArray + indx + start_i, 0, sizeof(UChar)*rwidth); 1636 } 1637 1638 return ; 1639 } 1640 1641 /************************************************************* 1642 Function: MoveNeighborSAD 1643 Date: 3/27/01 1644 Purpose: Move neighboring SAD around when center has shifted 1645 *************************************************************/ 1646 1647 void MoveNeighborSAD(Int dn[], Int new_loc) 1648 { 1649 Int tmp[9]; 1650 tmp[0] = dn[0]; 1651 tmp[1] = dn[1]; 1652 tmp[2] = dn[2]; 1653 tmp[3] = dn[3]; 1654 tmp[4] = dn[4]; 1655 tmp[5] = dn[5]; 1656 tmp[6] = dn[6]; 1657 tmp[7] = dn[7]; 1658 tmp[8] = dn[8]; 1659 dn[0] = dn[1] = dn[2] = dn[3] = dn[4] = dn[5] = dn[6] = dn[7] = dn[8] = 65536; 1660 1661 switch (new_loc) 1662 { 1663 case 0: 1664 break; 1665 case 1: 1666 dn[4] = tmp[2]; 1667 dn[5] = tmp[0]; 1668 dn[6] = tmp[8]; 1669 break; 1670 case 2: 1671 dn[4] = tmp[3]; 1672 dn[5] = tmp[4]; 1673 dn[6] = tmp[0]; 1674 dn[7] = tmp[8]; 1675 dn[8] = tmp[1]; 1676 break; 1677 case 3: 1678 dn[6] = tmp[4]; 1679 dn[7] = tmp[0]; 1680 dn[8] = tmp[2]; 1681 break; 1682 case 4: 1683 dn[1] = tmp[2]; 1684 dn[2] = tmp[3]; 1685 dn[6] = tmp[5]; 1686 dn[7] = tmp[6]; 1687 dn[8] = tmp[0]; 1688 break; 1689 case 5: 1690 dn[1] = tmp[0]; 1691 dn[2] = tmp[4]; 1692 dn[8] = tmp[6]; 1693 break; 1694 case 6: 1695 dn[1] = tmp[8]; 1696 dn[2] = tmp[0]; 1697 dn[3] = tmp[4]; 1698 dn[4] = tmp[5]; 1699 dn[8] = tmp[7]; 1700 break; 1701 case 7: 1702 dn[2] = tmp[8]; 1703 dn[3] = tmp[0]; 1704 dn[4] = tmp[6]; 1705 break; 1706 case 8: 1707 dn[2] = tmp[1]; 1708 dn[3] = tmp[2]; 1709 dn[4] = tmp[0]; 1710 dn[5] = tmp[6]; 1711 dn[6] = tmp[7]; 1712 break; 1713 } 1714 dn[0] = tmp[new_loc]; 1715 1716 return ; 1717 } 1718 1719 /* 3/28/01, find minimal of dn[9] */ 1720 1721 Int FindMin(Int dn[]) 1722 { 1723 Int min, i; 1724 Int dmin; 1725 1726 dmin = dn[1]; 1727 min = 1; 1728 for (i = 2; i < 9; i++) 1729 { 1730 if (dn[i] < dmin) 1731 { 1732 dmin = dn[i]; 1733 min = i; 1734 } 1735 } 1736 1737 return min; 1738 } 1739 1740 1741 1742
