| 1 |
/************************************************************************** |
/***************************************************************************** |
| 2 |
* |
* |
| 3 |
* XVID MPEG-4 VIDEO CODEC |
* XVID MPEG-4 VIDEO CODEC |
| 4 |
* motion estimation |
* - Motion Estimation module - |
| 5 |
|
* |
| 6 |
|
* Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
| 7 |
|
* 2002 Michael Militzer <michael@xvid.org> |
| 8 |
* |
* |
| 9 |
* This program is an implementation of a part of one or more MPEG-4 |
* This program is an implementation of a part of one or more MPEG-4 |
| 10 |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
| 27 |
* |
* |
| 28 |
* You should have received a copy of the GNU General Public License |
* You should have received a copy of the GNU General Public License |
| 29 |
* along with this program; if not, write to the Free Software |
* along with this program; if not, write to the Free Software |
| 30 |
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
| 31 |
* |
* |
| 32 |
*************************************************************************/ |
*************************************************************************/ |
| 33 |
|
|
|
/************************************************************************** |
|
|
* |
|
|
* Modifications: |
|
|
* |
|
|
* 01.05.2002 updated MotionEstimationBVOP |
|
|
* 25.04.2002 partial prevMB conversion |
|
|
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
|
|
* 14.04.2002 added MotionEstimationBVOP() |
|
|
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
|
|
* EPZS and EPZS^2 |
|
|
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
|
|
* PMVFast_Refine to support multiple searches with different start points |
|
|
* 07.01.2002 uv-block-based interpolation |
|
|
* 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
|
|
* changed INTER_BIAS to 150 (as suggested by suxen_drol) |
|
|
* removed halfpel refinement step in PMVfastSearch8 + quality=5 |
|
|
* added new quality mode = 6 which performs halfpel refinement |
|
|
* filesize difference between quality 5 and 6 is smaller than 1% |
|
|
* (Isibaar) |
|
|
* 31.12.2001 PMVfastSearch16 and PMVfastSearch8 (gruel) |
|
|
* 30.12.2001 get_range/MotionSearchX simplified; blue/green bug fix |
|
|
* 22.12.2001 commented best_point==99 check |
|
|
* 19.12.2001 modified get_range (purple bug fix) |
|
|
* 15.12.2001 moved pmv displacement from mbprediction |
|
|
* 02.12.2001 motion estimation/compensation split (Isibaar) |
|
|
* 16.11.2001 rewrote/tweaked search algorithms; pross@cs.rmit.edu.au |
|
|
* 10.11.2001 support for sad16/sad8 functions |
|
|
* 28.08.2001 reactivated MODE_INTER4V for EXT_MODE |
|
|
* 24.08.2001 removed MODE_INTER4V_Q, disabled MODE_INTER4V for EXT_MODE |
|
|
* 22.08.2001 added MODE_INTER4V_Q |
|
|
* 20.08.2001 added pragma to get rid of internal compiler error with VC6 |
|
|
* idea by Cyril. Thanks. |
|
|
* |
|
|
* Michael Militzer <isibaar@videocoding.de> |
|
|
* |
|
|
**************************************************************************/ |
|
|
|
|
| 34 |
#include <assert.h> |
#include <assert.h> |
| 35 |
#include <stdio.h> |
#include <stdio.h> |
| 36 |
#include <stdlib.h> |
#include <stdlib.h> |
| 140 |
static const VECTOR zeroMV = { 0, 0 }; |
static const VECTOR zeroMV = { 0, 0 }; |
| 141 |
VECTOR predMV; |
VECTOR predMV; |
| 142 |
|
|
| 143 |
int32_t x, y; |
uint32_t x, y; |
| 144 |
int32_t iIntra = 0; |
uint32_t iIntra = 0; |
| 145 |
VECTOR pmv; |
VECTOR pmv; |
| 146 |
|
|
| 147 |
if (sadInit) |
if (sadInit) |
| 355 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
| 356 |
} |
} |
| 357 |
|
|
| 358 |
|
#if 0 |
| 359 |
/* too slow and not fully functional at the moment */ |
/* too slow and not fully functional at the moment */ |
|
/* |
|
| 360 |
int32_t ZeroSearch16( |
int32_t ZeroSearch16( |
| 361 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
| 362 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
| 395 |
return iSAD; |
return iSAD; |
| 396 |
|
|
| 397 |
} |
} |
| 398 |
*/ |
#endif /* 0 */ |
| 399 |
|
|
| 400 |
int32_t |
int32_t |
| 401 |
Diamond16_MainSearch(const uint8_t * const pRef, |
Diamond16_MainSearch(const uint8_t * const pRef, |
| 424 |
/* Do a diamond search around given starting point, return SAD of best */ |
/* Do a diamond search around given starting point, return SAD of best */ |
| 425 |
|
|
| 426 |
int32_t iDirection = 0; |
int32_t iDirection = 0; |
| 427 |
|
int32_t iDirectionBackup; |
| 428 |
int32_t iSAD; |
int32_t iSAD; |
| 429 |
VECTOR backupMV; |
VECTOR backupMV; |
| 430 |
|
|
| 438 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); |
| 439 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); |
| 440 |
|
|
| 441 |
if (iDirection) |
if (iDirection) { |
| 442 |
while (!iFound) { |
while (!iFound) { |
| 443 |
iFound = 1; |
iFound = 1; |
| 444 |
backupMV = *currMV; |
backupMV = *currMV; |
| 445 |
|
iDirectionBackup = iDirection; |
| 446 |
|
|
| 447 |
if (iDirection != 2) |
if (iDirectionBackup != 2) |
| 448 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 449 |
backupMV.y, 1); |
backupMV.y, 1); |
| 450 |
if (iDirection != 1) |
if (iDirectionBackup != 1) |
| 451 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 452 |
backupMV.y, 2); |
backupMV.y, 2); |
| 453 |
if (iDirection != 4) |
if (iDirectionBackup != 4) |
| 454 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, |
| 455 |
backupMV.y - iDiamondSize, 3); |
backupMV.y - iDiamondSize, 3); |
| 456 |
if (iDirection != 3) |
if (iDirectionBackup != 3) |
| 457 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, |
| 458 |
backupMV.y + iDiamondSize, 4); |
backupMV.y + iDiamondSize, 4); |
| 459 |
|
} |
| 460 |
} else { |
} else { |
| 461 |
currMV->x = start_x; |
currMV->x = start_x; |
| 462 |
currMV->y = start_y; |
currMV->y = start_y; |
| 520 |
backupMV.y + iDiamondSize, 8); |
backupMV.y + iDiamondSize, 8); |
| 521 |
|
|
| 522 |
|
|
| 523 |
if (iDirection) |
if (iDirection) { |
| 524 |
while (!iFound) { |
while (!iFound) { |
| 525 |
iFound = 1; |
iFound = 1; |
| 526 |
backupMV = *currMV; |
backupMV = *currMV; |
| 529 |
case 1: |
case 1: |
| 530 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 531 |
backupMV.y, 1); |
backupMV.y, 1); |
| 532 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 533 |
backupMV.y - iDiamondSize, 5); |
backupMV.y - iDiamondSize, 5); |
| 534 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 535 |
backupMV.y - iDiamondSize, 7); |
backupMV.y - iDiamondSize, 7); |
| 536 |
break; |
break; |
| 537 |
case 2: |
case 2: |
| 538 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
| 539 |
2); |
2); |
| 540 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 541 |
backupMV.y + iDiamondSize, 6); |
backupMV.y + iDiamondSize, 6); |
| 542 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 543 |
backupMV.y + iDiamondSize, 8); |
backupMV.y + iDiamondSize, 8); |
| 544 |
break; |
break; |
| 545 |
|
|
| 546 |
case 3: |
case 3: |
| 547 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
| 548 |
4); |
4); |
| 549 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 550 |
backupMV.y - iDiamondSize, 7); |
backupMV.y - iDiamondSize, 7); |
| 551 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 552 |
backupMV.y + iDiamondSize, 8); |
backupMV.y + iDiamondSize, 8); |
| 553 |
break; |
break; |
| 554 |
|
|
| 555 |
case 4: |
case 4: |
| 556 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
| 557 |
3); |
3); |
| 558 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 559 |
backupMV.y - iDiamondSize, 5); |
backupMV.y - iDiamondSize, 5); |
| 560 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 561 |
backupMV.y + iDiamondSize, 6); |
backupMV.y + iDiamondSize, 6); |
| 562 |
break; |
break; |
| 563 |
|
|
| 564 |
case 5: |
case 5: |
| 565 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, |
| 566 |
1); |
1); |
| 567 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
| 568 |
3); |
3); |
| 569 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 570 |
backupMV.y - iDiamondSize, 5); |
backupMV.y - iDiamondSize, 5); |
| 571 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 572 |
backupMV.y + iDiamondSize, 6); |
backupMV.y + iDiamondSize, 6); |
| 573 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 574 |
backupMV.y - iDiamondSize, 7); |
backupMV.y - iDiamondSize, 7); |
| 575 |
break; |
break; |
| 576 |
|
|
| 577 |
case 6: |
case 6: |
| 578 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
| 579 |
2); |
2); |
| 580 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
| 581 |
3); |
3); |
| 582 |
|
|
| 583 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 584 |
backupMV.y - iDiamondSize, 5); |
backupMV.y - iDiamondSize, 5); |
| 585 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 586 |
backupMV.y + iDiamondSize, 6); |
backupMV.y + iDiamondSize, 6); |
| 587 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 588 |
backupMV.y + iDiamondSize, 8); |
backupMV.y + iDiamondSize, 8); |
| 589 |
|
|
| 590 |
break; |
break; |
| 592 |
case 7: |
case 7: |
| 593 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 594 |
backupMV.y, 1); |
backupMV.y, 1); |
| 595 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
| 596 |
4); |
4); |
| 597 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 598 |
backupMV.y - iDiamondSize, 5); |
backupMV.y - iDiamondSize, 5); |
| 599 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 600 |
backupMV.y - iDiamondSize, 7); |
backupMV.y - iDiamondSize, 7); |
| 601 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 602 |
backupMV.y + iDiamondSize, 8); |
backupMV.y + iDiamondSize, 8); |
| 603 |
break; |
break; |
| 604 |
|
|
| 605 |
case 8: |
case 8: |
| 606 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
| 607 |
2); |
2); |
| 608 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
| 609 |
4); |
4); |
| 610 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 611 |
backupMV.y + iDiamondSize, 6); |
backupMV.y + iDiamondSize, 6); |
| 612 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 613 |
backupMV.y - iDiamondSize, 7); |
backupMV.y - iDiamondSize, 7); |
| 614 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 615 |
backupMV.y + iDiamondSize, 8); |
backupMV.y + iDiamondSize, 8); |
| 616 |
break; |
break; |
| 617 |
default: |
default: |
| 618 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, |
| 619 |
1); |
1); |
| 620 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
| 621 |
2); |
2); |
| 622 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
| 623 |
3); |
3); |
| 624 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
| 625 |
4); |
4); |
| 626 |
|
|
| 627 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 628 |
backupMV.y - iDiamondSize, 5); |
backupMV.y - iDiamondSize, 5); |
| 629 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 630 |
backupMV.y + iDiamondSize, 6); |
backupMV.y + iDiamondSize, 6); |
| 631 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 632 |
backupMV.y - iDiamondSize, 7); |
backupMV.y - iDiamondSize, 7); |
| 633 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 634 |
backupMV.y + iDiamondSize, 8); |
backupMV.y + iDiamondSize, 8); |
| 635 |
break; |
break; |
| 636 |
} |
} |
| 637 |
|
} |
| 638 |
} else { |
} else { |
| 639 |
currMV->x = start_x; |
currMV->x = start_x; |
| 640 |
currMV->y = start_y; |
currMV->y = start_y; |
| 689 |
const uint8_t * const cur, |
const uint8_t * const cur, |
| 690 |
const int x, |
const int x, |
| 691 |
const int y, |
const int y, |
| 692 |
int start_x, |
const int start_xi, |
| 693 |
int start_y, |
const int start_yi, |
| 694 |
int iMinSAD, |
int iMinSAD, |
| 695 |
VECTOR * const currMV, |
VECTOR * const currMV, |
| 696 |
const int center_x, |
const int center_x, |
| 707 |
{ |
{ |
| 708 |
|
|
| 709 |
int32_t iSAD; |
int32_t iSAD; |
| 710 |
|
int start_x = start_xi, start_y = start_yi; |
| 711 |
|
|
| 712 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
| 713 |
|
|
| 838 |
} |
} |
| 839 |
while (1); //forever |
while (1); //forever |
| 840 |
} |
} |
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
#define CHECK_MV16_F_INTERPOL(X,Y,BX,BY) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - center_x, (Y) - center_y, (uint8_t)iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_F_INTERPOL_DIR(X,Y,BX,BY,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - center_x, (Y) - center_y, (uint8_t)iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_F_INTERPOL_FOUND(X,Y,BX,BY,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - center_x, (Y) - center_y, (uint8_t)iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
|
} |
|
|
|
|
|
|
|
|
#define CHECK_MV16_B_INTERPOL(FX,FY,X,Y) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - center_x, (Y) - center_y, (uint8_t)iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
|
|
|
#define CHECK_MV16_B_INTERPOL_DIR(FX,FY,X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - center_x, (Y) - center_y, (uint8_t)iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
|
|
} |
|
|
|
|
|
|
|
|
#define CHECK_MV16_B_INTERPOL_FOUND(FX,FY,X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - center_x, (Y) - center_y, (uint8_t)iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
|
} |
|
|
|
|
|
|
|
|
#if (0==1) |
|
|
int32_t |
|
|
Diamond16_InterpolMainSearch( |
|
|
const uint8_t * const f_pRef, |
|
|
const uint8_t * const f_pRefH, |
|
|
const uint8_t * const f_pRefV, |
|
|
const uint8_t * const f_pRefHV, |
|
|
const uint8_t * const cur, |
|
|
|
|
|
const uint8_t * const b_pRef, |
|
|
const uint8_t * const b_pRefH, |
|
|
const uint8_t * const b_pRefV, |
|
|
const uint8_t * const b_pRefHV, |
|
|
|
|
|
const int x, |
|
|
const int y, |
|
| 841 |
|
|
|
const int f_start_x, |
|
|
const int f_start_y, |
|
|
const int b_start_x, |
|
|
const int b_start_y, |
|
|
|
|
|
int iMinSAD, |
|
|
VECTOR * const f_currMV, |
|
|
VECTOR * const b_currMV, |
|
|
|
|
|
const int f_center_x, |
|
|
const int f_center_y, |
|
|
const int b_center_x, |
|
|
const int b_center_y, |
|
|
|
|
|
const int32_t min_dx, |
|
|
const int32_t max_dx, |
|
|
const int32_t min_dy, |
|
|
const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
|
|
|
const int32_t f_iFcode, |
|
|
const int32_t b_iFcode, |
|
|
|
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t f_iDirection = 0; |
|
|
int32_t b_iDirection = 0; |
|
|
int32_t iSAD; |
|
|
|
|
|
VECTOR f_backupMV; |
|
|
VECTOR b_backupMV; |
|
|
|
|
|
f_backupMV.x = start_x; |
|
|
f_backupMV.y = start_y; |
|
|
b_backupMV.x = start_x; |
|
|
b_backupMV.y = start_y; |
|
|
|
|
|
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
|
|
|
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, 1); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, 2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); |
|
|
|
|
|
if (iDirection) |
|
|
while (!iFound) { |
|
|
iFound = 1; |
|
|
backupMV = *currMV; |
|
|
|
|
|
if (iDirection != 2) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y, 1); |
|
|
if (iDirection != 1) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y, 2); |
|
|
if (iDirection != 4) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, |
|
|
backupMV.y - iDiamondSize, 3); |
|
|
if (iDirection != 3) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, |
|
|
backupMV.y + iDiamondSize, 4); |
|
|
} else { |
|
|
currMV->x = start_x; |
|
|
currMV->y = start_y; |
|
|
} |
|
| 842 |
return iMinSAD; |
return iMinSAD; |
| 843 |
} |
} |
|
#endif |
|
|
|
|
| 844 |
|
|
| 845 |
int32_t |
int32_t |
| 846 |
AdvDiamond8_MainSearch(const uint8_t * const pRef, |
AdvDiamond8_MainSearch(const uint8_t * const pRef, |
| 850 |
const uint8_t * const cur, |
const uint8_t * const cur, |
| 851 |
const int x, |
const int x, |
| 852 |
const int y, |
const int y, |
| 853 |
int start_x, |
const int start_xi, |
| 854 |
int start_y, |
const int start_yi, |
| 855 |
int iMinSAD, |
int iMinSAD, |
| 856 |
VECTOR * const currMV, |
VECTOR * const currMV, |
| 857 |
const int center_x, |
const int center_x, |
| 868 |
{ |
{ |
| 869 |
|
|
| 870 |
int32_t iSAD; |
int32_t iSAD; |
| 871 |
|
int start_x = start_xi, start_y = start_yi; |
| 872 |
|
|
| 873 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
| 874 |
|
|
| 1091 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
| 1092 |
const int x, |
const int x, |
| 1093 |
const int y, |
const int y, |
| 1094 |
const int start_x, |
const int start_x, /* start is searched first, so it should contain the most */ |
| 1095 |
const int start_y, |
const int start_y, /* likely motion vector for this block */ |
| 1096 |
const int center_x, |
const int center_x, /* center is from where length of MVs is measured */ |
| 1097 |
const int center_y, |
const int center_y, |
| 1098 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
| 1099 |
const uint32_t iQuant, |
const uint32_t iQuant, |
| 1148 |
} |
} |
| 1149 |
|
|
| 1150 |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
//bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
|
| 1151 |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
| 1152 |
|
|
| 1153 |
if ((x == 0) && (y == 0)) { |
if ((x == 0) && (y == 0)) { |
| 1205 |
if ((iMinSAD < 256) || |
if ((iMinSAD < 256) || |
| 1206 |
((MVequal(*currMV, prevMB->mvs[0])) && |
((MVequal(*currMV, prevMB->mvs[0])) && |
| 1207 |
((int32_t) iMinSAD < prevMB->sad16))) { |
((int32_t) iMinSAD < prevMB->sad16))) { |
| 1208 |
if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode |
if (iMinSAD < (int)(2 * iQuant)) // high chances for SKIP-mode |
| 1209 |
{ |
{ |
| 1210 |
if (!MVzero(*currMV)) { |
if (!MVzero(*currMV)) { |
| 1211 |
iMinSAD += MV16_00_BIAS; |
iMinSAD += MV16_00_BIAS; |
| 1410 |
const uint8_t * const cur, |
const uint8_t * const cur, |
| 1411 |
const int x, |
const int x, |
| 1412 |
const int y, |
const int y, |
| 1413 |
int32_t start_x, |
const int32_t start_x, |
| 1414 |
int32_t start_y, |
const int32_t start_y, |
| 1415 |
int32_t iMinSAD, |
int32_t iMinSAD, |
| 1416 |
VECTOR * const currMV, |
VECTOR * const currMV, |
| 1417 |
const int center_x, |
const int center_x, |
| 1429 |
/* Do a diamond search around given starting point, return SAD of best */ |
/* Do a diamond search around given starting point, return SAD of best */ |
| 1430 |
|
|
| 1431 |
int32_t iDirection = 0; |
int32_t iDirection = 0; |
| 1432 |
|
int32_t iDirectionBackup; |
| 1433 |
int32_t iSAD; |
int32_t iSAD; |
| 1434 |
VECTOR backupMV; |
VECTOR backupMV; |
| 1435 |
|
|
| 1443 |
CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); |
CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); |
| 1444 |
CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); |
CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); |
| 1445 |
|
|
| 1446 |
if (iDirection) |
if (iDirection) { |
| 1447 |
while (!iFound) { |
while (!iFound) { |
| 1448 |
iFound = 1; |
iFound = 1; |
| 1449 |
backupMV = *currMV; // since iDirection!=0, this is well defined! |
backupMV = *currMV; // since iDirection!=0, this is well defined! |
| 1450 |
|
iDirectionBackup = iDirection; |
| 1451 |
|
|
| 1452 |
if (iDirection != 2) |
if (iDirectionBackup != 2) |
| 1453 |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1454 |
backupMV.y, 1); |
backupMV.y, 1); |
| 1455 |
if (iDirection != 1) |
if (iDirectionBackup != 1) |
| 1456 |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1457 |
backupMV.y, 2); |
backupMV.y, 2); |
| 1458 |
if (iDirection != 4) |
if (iDirectionBackup != 4) |
| 1459 |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, |
| 1460 |
backupMV.y - iDiamondSize, 3); |
backupMV.y - iDiamondSize, 3); |
| 1461 |
if (iDirection != 3) |
if (iDirectionBackup != 3) |
| 1462 |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, |
| 1463 |
backupMV.y + iDiamondSize, 4); |
backupMV.y + iDiamondSize, 4); |
| 1464 |
|
} |
| 1465 |
} else { |
} else { |
| 1466 |
currMV->x = start_x; |
currMV->x = start_x; |
| 1467 |
currMV->y = start_y; |
currMV->y = start_y; |
| 1469 |
return iMinSAD; |
return iMinSAD; |
| 1470 |
} |
} |
| 1471 |
|
|
| 1472 |
|
|
| 1473 |
|
|
| 1474 |
|
|
| 1475 |
int32_t |
int32_t |
| 1476 |
Halfpel8_Refine_c(const uint8_t * const pRef, |
Square8_MainSearch(const uint8_t * const pRef, |
| 1477 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
| 1478 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
| 1479 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
| 1480 |
const uint8_t * const cur, |
const uint8_t * const cur, |
| 1481 |
const int x, |
const int x, |
| 1482 |
const int y, |
const int y, |
| 1483 |
VECTOR * const currMV, |
const int32_t start_x, |
| 1484 |
|
const int32_t start_y, |
| 1485 |
int32_t iMinSAD, |
int32_t iMinSAD, |
| 1486 |
|
VECTOR * const currMV, |
| 1487 |
const int center_x, |
const int center_x, |
| 1488 |
const int center_y, |
const int center_y, |
| 1489 |
const int32_t min_dx, |
const int32_t min_dx, |
| 1490 |
const int32_t max_dx, |
const int32_t max_dx, |
| 1491 |
const int32_t min_dy, |
const int32_t min_dy, |
| 1492 |
const int32_t max_dy, |
const int32_t max_dy, |
| 1493 |
|
const int32_t iEdgedWidth, |
| 1494 |
|
const int32_t iDiamondSize, |
| 1495 |
const int32_t iFcode, |
const int32_t iFcode, |
| 1496 |
const int32_t iQuant, |
const int32_t iQuant, |
| 1497 |
const int32_t iEdgedWidth) |
int iFound) |
| 1498 |
{ |
{ |
| 1499 |
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
/* Do a square search around given starting point, return SAD of best */ |
| 1500 |
|
|
| 1501 |
|
int32_t iDirection = 0; |
| 1502 |
int32_t iSAD; |
int32_t iSAD; |
| 1503 |
VECTOR backupMV = *currMV; |
VECTOR backupMV; |
|
|
|
|
CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y - 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x, backupMV.y - 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y - 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y + 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x, backupMV.y + 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y + 1); |
|
| 1504 |
|
|
| 1505 |
return iMinSAD; |
backupMV.x = start_x; |
| 1506 |
} |
backupMV.y = start_y; |
| 1507 |
|
|
| 1508 |
|
/* It's one search with full square pattern, and new parts for all following diamonds */ |
| 1509 |
|
|
| 1510 |
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
/* new direction are extra, so 1-4 is normal diamond |
| 1511 |
|
537 |
| 1512 |
|
1*2 |
| 1513 |
|
648 |
| 1514 |
|
*/ |
| 1515 |
|
|
| 1516 |
int32_t |
CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, 1); |
| 1517 |
PMVfastSearch8(const uint8_t * const pRef, |
CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, 2); |
| 1518 |
const uint8_t * const pRefH, |
CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); |
| 1519 |
const uint8_t * const pRefV, |
CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); |
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, |
|
|
const int y, |
|
|
const int start_x, |
|
|
const int start_y, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
| 1520 |
|
|
| 1521 |
const uint8_t *cur = pCur->y + x * 8 + y * 8 * iEdgedWidth; |
CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
| 1522 |
|
backupMV.y - iDiamondSize, 5); |
| 1523 |
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
| 1524 |
|
backupMV.y + iDiamondSize, 6); |
| 1525 |
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
| 1526 |
|
backupMV.y - iDiamondSize, 7); |
| 1527 |
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
| 1528 |
|
backupMV.y + iDiamondSize, 8); |
| 1529 |
|
|
|
int32_t iDiamondSize; |
|
| 1530 |
|
|
| 1531 |
int32_t min_dx; |
if (iDirection) { |
| 1532 |
int32_t max_dx; |
while (!iFound) { |
| 1533 |
int32_t min_dy; |
iFound = 1; |
| 1534 |
int32_t max_dy; |
backupMV = *currMV; |
| 1535 |
|
|
| 1536 |
|
switch (iDirection) { |
| 1537 |
|
case 1: |
| 1538 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1539 |
|
backupMV.y, 1); |
| 1540 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1541 |
|
backupMV.y - iDiamondSize, 5); |
| 1542 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1543 |
|
backupMV.y - iDiamondSize, 7); |
| 1544 |
|
break; |
| 1545 |
|
case 2: |
| 1546 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
| 1547 |
|
2); |
| 1548 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1549 |
|
backupMV.y + iDiamondSize, 6); |
| 1550 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1551 |
|
backupMV.y + iDiamondSize, 8); |
| 1552 |
|
break; |
| 1553 |
|
|
| 1554 |
|
case 3: |
| 1555 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
| 1556 |
|
4); |
| 1557 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1558 |
|
backupMV.y - iDiamondSize, 7); |
| 1559 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1560 |
|
backupMV.y + iDiamondSize, 8); |
| 1561 |
|
break; |
| 1562 |
|
|
| 1563 |
|
case 4: |
| 1564 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
| 1565 |
|
3); |
| 1566 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1567 |
|
backupMV.y - iDiamondSize, 5); |
| 1568 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1569 |
|
backupMV.y + iDiamondSize, 6); |
| 1570 |
|
break; |
| 1571 |
|
|
| 1572 |
|
case 5: |
| 1573 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, |
| 1574 |
|
1); |
| 1575 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
| 1576 |
|
3); |
| 1577 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1578 |
|
backupMV.y - iDiamondSize, 5); |
| 1579 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1580 |
|
backupMV.y + iDiamondSize, 6); |
| 1581 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1582 |
|
backupMV.y - iDiamondSize, 7); |
| 1583 |
|
break; |
| 1584 |
|
|
| 1585 |
|
case 6: |
| 1586 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
| 1587 |
|
2); |
| 1588 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
| 1589 |
|
3); |
| 1590 |
|
|
| 1591 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1592 |
|
backupMV.y - iDiamondSize, 5); |
| 1593 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1594 |
|
backupMV.y + iDiamondSize, 6); |
| 1595 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1596 |
|
backupMV.y + iDiamondSize, 8); |
| 1597 |
|
|
| 1598 |
|
break; |
| 1599 |
|
|
| 1600 |
|
case 7: |
| 1601 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1602 |
|
backupMV.y, 1); |
| 1603 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
| 1604 |
|
4); |
| 1605 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1606 |
|
backupMV.y - iDiamondSize, 5); |
| 1607 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1608 |
|
backupMV.y - iDiamondSize, 7); |
| 1609 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1610 |
|
backupMV.y + iDiamondSize, 8); |
| 1611 |
|
break; |
| 1612 |
|
|
| 1613 |
|
case 8: |
| 1614 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
| 1615 |
|
2); |
| 1616 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
| 1617 |
|
4); |
| 1618 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1619 |
|
backupMV.y + iDiamondSize, 6); |
| 1620 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1621 |
|
backupMV.y - iDiamondSize, 7); |
| 1622 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1623 |
|
backupMV.y + iDiamondSize, 8); |
| 1624 |
|
break; |
| 1625 |
|
default: |
| 1626 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, |
| 1627 |
|
1); |
| 1628 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
| 1629 |
|
2); |
| 1630 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
| 1631 |
|
3); |
| 1632 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
| 1633 |
|
4); |
| 1634 |
|
|
| 1635 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1636 |
|
backupMV.y - iDiamondSize, 5); |
| 1637 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
| 1638 |
|
backupMV.y + iDiamondSize, 6); |
| 1639 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1640 |
|
backupMV.y - iDiamondSize, 7); |
| 1641 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
| 1642 |
|
backupMV.y + iDiamondSize, 8); |
| 1643 |
|
break; |
| 1644 |
|
} |
| 1645 |
|
} |
| 1646 |
|
} else { |
| 1647 |
|
currMV->x = start_x; |
| 1648 |
|
currMV->y = start_y; |
| 1649 |
|
} |
| 1650 |
|
return iMinSAD; |
| 1651 |
|
} |
| 1652 |
|
|
| 1653 |
|
|
| 1654 |
|
|
| 1655 |
|
|
| 1656 |
|
|
| 1657 |
|
int32_t |
| 1658 |
|
Halfpel8_Refine_c(const uint8_t * const pRef, |
| 1659 |
|
const uint8_t * const pRefH, |
| 1660 |
|
const uint8_t * const pRefV, |
| 1661 |
|
const uint8_t * const pRefHV, |
| 1662 |
|
const uint8_t * const cur, |
| 1663 |
|
const int x, |
| 1664 |
|
const int y, |
| 1665 |
|
VECTOR * const currMV, |
| 1666 |
|
int32_t iMinSAD, |
| 1667 |
|
const int center_x, |
| 1668 |
|
const int center_y, |
| 1669 |
|
const int32_t min_dx, |
| 1670 |
|
const int32_t max_dx, |
| 1671 |
|
const int32_t min_dy, |
| 1672 |
|
const int32_t max_dy, |
| 1673 |
|
const int32_t iFcode, |
| 1674 |
|
const int32_t iQuant, |
| 1675 |
|
const int32_t iEdgedWidth) |
| 1676 |
|
{ |
| 1677 |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
| 1678 |
|
|
| 1679 |
|
int32_t iSAD; |
| 1680 |
|
VECTOR backupMV = *currMV; |
| 1681 |
|
|
| 1682 |
|
CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y - 1); |
| 1683 |
|
CHECK_MV8_CANDIDATE(backupMV.x, backupMV.y - 1); |
| 1684 |
|
CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y - 1); |
| 1685 |
|
CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y); |
| 1686 |
|
CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y); |
| 1687 |
|
CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y + 1); |
| 1688 |
|
CHECK_MV8_CANDIDATE(backupMV.x, backupMV.y + 1); |
| 1689 |
|
CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y + 1); |
| 1690 |
|
|
| 1691 |
|
return iMinSAD; |
| 1692 |
|
} |
| 1693 |
|
|
| 1694 |
|
|
| 1695 |
|
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
| 1696 |
|
|
| 1697 |
|
int32_t |
| 1698 |
|
PMVfastSearch8(const uint8_t * const pRef, |
| 1699 |
|
const uint8_t * const pRefH, |
| 1700 |
|
const uint8_t * const pRefV, |
| 1701 |
|
const uint8_t * const pRefHV, |
| 1702 |
|
const IMAGE * const pCur, |
| 1703 |
|
const int x, |
| 1704 |
|
const int y, |
| 1705 |
|
const int start_x, |
| 1706 |
|
const int start_y, |
| 1707 |
|
const int center_x, |
| 1708 |
|
const int center_y, |
| 1709 |
|
const uint32_t MotionFlags, |
| 1710 |
|
const uint32_t iQuant, |
| 1711 |
|
const uint32_t iFcode, |
| 1712 |
|
const MBParam * const pParam, |
| 1713 |
|
const MACROBLOCK * const pMBs, |
| 1714 |
|
const MACROBLOCK * const prevMBs, |
| 1715 |
|
VECTOR * const currMV, |
| 1716 |
|
VECTOR * const currPMV) |
| 1717 |
|
{ |
| 1718 |
|
const uint32_t iWcount = pParam->mb_width; |
| 1719 |
|
const int32_t iWidth = pParam->width; |
| 1720 |
|
const int32_t iHeight = pParam->height; |
| 1721 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
| 1722 |
|
|
| 1723 |
|
const uint8_t *cur = pCur->y + x * 8 + y * 8 * iEdgedWidth; |
| 1724 |
|
|
| 1725 |
|
int32_t iDiamondSize; |
| 1726 |
|
|
| 1727 |
|
int32_t min_dx; |
| 1728 |
|
int32_t max_dx; |
| 1729 |
|
int32_t min_dy; |
| 1730 |
|
int32_t max_dy; |
| 1731 |
|
|
| 1732 |
VECTOR pmv[4]; |
VECTOR pmv[4]; |
| 1733 |
int32_t psad[4]; |
int32_t psad[4]; |
| 1762 |
} |
} |
| 1763 |
|
|
| 1764 |
/* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
/* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
|
//bPredEq = get_pmvdata(pMBs, (x >> 1), (y >> 1), iWcount, iSubBlock, pmv, psad); |
|
| 1765 |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, (x >> 1), (y >> 1), iSubBlock, pmv, psad); |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, (x >> 1), (y >> 1), iSubBlock, pmv, psad); |
| 1766 |
|
|
| 1767 |
if ((x == 0) && (y == 0)) { |
if ((x == 0) && (y == 0)) { |
| 1791 |
|
|
| 1792 |
// Prepare for main loop |
// Prepare for main loop |
| 1793 |
|
|
| 1794 |
// if (MotionFlags & PMV_USESQUARES8) |
if (MotionFlags & PMV_USESQUARES8) |
| 1795 |
// MainSearchPtr = Square8_MainSearch; |
MainSearchPtr = Square8_MainSearch; |
| 1796 |
// else |
else |
| 1797 |
|
|
| 1798 |
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
| 1799 |
MainSearchPtr = AdvDiamond8_MainSearch; |
MainSearchPtr = AdvDiamond8_MainSearch; |
| 2076 |
max_dy = EVEN(max_dy); |
max_dy = EVEN(max_dy); |
| 2077 |
} |
} |
| 2078 |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
//bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
|
| 2079 |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
| 2080 |
|
|
| 2081 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
| 2358 |
max_dy = EVEN(max_dy); |
max_dy = EVEN(max_dy); |
| 2359 |
} |
} |
| 2360 |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
//bPredEq = get_pmvdata(pMBs, x >> 1, y >> 1, iWcount, iSubBlock, pmv[0].x, pmv[0].y, psad); |
|
| 2361 |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x >> 1, y >> 1, iSubBlock, pmv, psad); |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x >> 1, y >> 1, iSubBlock, pmv, psad); |
| 2362 |
|
|
| 2363 |
|
|
| 2469 |
|
|
| 2470 |
// there is no EPZS^2 for inter4v at the moment |
// there is no EPZS^2 for inter4v at the moment |
| 2471 |
|
|
| 2472 |
// if (MotionFlags & PMV_USESQUARES8) |
if (MotionFlags & PMV_USESQUARES8) |
| 2473 |
// MainSearchPtr = Square8_MainSearch; |
MainSearchPtr = Square8_MainSearch; |
| 2474 |
// else |
else |
| 2475 |
|
|
| 2476 |
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
| 2477 |
MainSearchPtr = AdvDiamond8_MainSearch; |
MainSearchPtr = AdvDiamond8_MainSearch; |
| 2535 |
currPMV->y = currMV->y - center_y; |
currPMV->y = currMV->y - center_y; |
| 2536 |
return iMinSAD; |
return iMinSAD; |
| 2537 |
} |
} |
|
|
|
|
|
|
|
|
|
|
int32_t |
|
|
PMVfastIntSearch16(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, |
|
|
const int y, |
|
|
const int start_x, |
|
|
const int start_y, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; |
|
|
const VECTOR zeroMV = { 0, 0 }; |
|
|
|
|
|
int32_t iDiamondSize; |
|
|
|
|
|
int32_t min_dx; |
|
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
|
|
|
|
int32_t iFound; |
|
|
|
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; /* just for PMVFAST */ |
|
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[4]; |
|
|
|
|
|
MainSearch16FuncPtr MainSearchPtr; |
|
|
|
|
|
const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; |
|
|
MACROBLOCK *const pMB = pMBs + x + y * iWcount; |
|
|
|
|
|
int32_t threshA, threshB; |
|
|
int32_t bPredEq; |
|
|
int32_t iMinSAD, iSAD; |
|
|
|
|
|
|
|
|
/* Get maximum range */ |
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, |
|
|
iFcode); |
|
|
|
|
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
|
|
|
|
|
if ((x == 0) && (y == 0)) { |
|
|
threshA = 512; |
|
|
threshB = 1024; |
|
|
|
|
|
bPredEq = 0; |
|
|
psad[0] = psad[1] = psad[2] = psad[3] = 0; |
|
|
*currMV = pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV; |
|
|
|
|
|
} else { |
|
|
threshA = psad[0]; |
|
|
threshB = threshA + 256; |
|
|
if (threshA < 512) |
|
|
threshA = 512; |
|
|
if (threshA > 1024) |
|
|
threshA = 1024; |
|
|
if (threshB > 1792) |
|
|
threshB = 1792; |
|
|
|
|
|
bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
|
|
*currMV = pmv[0]; /* current best := prediction */ |
|
|
} |
|
|
|
|
|
iFound = 0; |
|
|
|
|
|
/* Step 4: Calculate SAD around the Median prediction. |
|
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
|
|
|
|
if (currMV->x > max_dx) { |
|
|
currMV->x = EVEN(max_dx); |
|
|
} |
|
|
if (currMV->x < min_dx) { |
|
|
currMV->x = EVEN(min_dx); |
|
|
} |
|
|
if (currMV->y > max_dy) { |
|
|
currMV->y = EVEN(max_dy); |
|
|
} |
|
|
if (currMV->y < min_dy) { |
|
|
currMV->y = EVEN(min_dy); |
|
|
} |
|
|
|
|
|
iMinSAD = |
|
|
sad16(cur, |
|
|
get_iref_mv(pRef, x, y, 16, currMV, |
|
|
iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += |
|
|
calc_delta_16(currMV->x - center_x, currMV->y - center_y, |
|
|
(uint8_t) iFcode, iQuant); |
|
|
|
|
|
if ((iMinSAD < 256) || |
|
|
((MVequal(*currMV, prevMB->i_mvs[0])) && |
|
|
((int32_t) iMinSAD < prevMB->i_sad16))) { |
|
|
if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode |
|
|
{ |
|
|
if (!MVzero(*currMV)) { |
|
|
iMinSAD += MV16_00_BIAS; |
|
|
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
|
|
iMinSAD -= MV16_00_BIAS; |
|
|
} |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfastInt16_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
|
|
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0], prevMB->i_mvs[0]))) |
|
|
iFound = 2; |
|
|
|
|
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) |
|
|
iDiamondSize = 2; // halfpel units! |
|
|
else |
|
|
iDiamondSize = 4; // halfpel units! |
|
|
|
|
|
/* |
|
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
|
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. |
|
|
*/ |
|
|
|
|
|
// (0,0) is often a good choice |
|
|
|
|
|
if (!MVzero(pmv[0])) |
|
|
CHECK_MV16_ZERO; |
|
|
|
|
|
// previous frame MV is always possible |
|
|
|
|
|
if (!MVzero(prevMB->i_mvs[0])) |
|
|
if (!MVequal(prevMB->i_mvs[0], pmv[0])) |
|
|
CHECK_MV16_CANDIDATE(prevMB->i_mvs[0].x, prevMB->i_mvs[0].y); |
|
|
|
|
|
// left neighbour, if allowed |
|
|
|
|
|
if (!MVzero(pmv[1])) |
|
|
if (!MVequal(pmv[1], prevMB->i_mvs[0])) |
|
|
if (!MVequal(pmv[1], pmv[0])) |
|
|
CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); |
|
|
|
|
|
// top neighbour, if allowed |
|
|
if (!MVzero(pmv[2])) |
|
|
if (!MVequal(pmv[2], prevMB->i_mvs[0])) |
|
|
if (!MVequal(pmv[2], pmv[0])) |
|
|
if (!MVequal(pmv[2], pmv[1])) |
|
|
CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); |
|
|
|
|
|
// top right neighbour, if allowed |
|
|
if (!MVzero(pmv[3])) |
|
|
if (!MVequal(pmv[3], prevMB->i_mvs[0])) |
|
|
if (!MVequal(pmv[3], pmv[0])) |
|
|
if (!MVequal(pmv[3], pmv[1])) |
|
|
if (!MVequal(pmv[3], pmv[2])) |
|
|
CHECK_MV16_CANDIDATE(pmv[3].x, |
|
|
pmv[3].y); |
|
|
|
|
|
if ((MVzero(*currMV)) && |
|
|
(!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
|
|
iMinSAD -= MV16_00_BIAS; |
|
|
|
|
|
|
|
|
/* Step 6: If MinSAD <= thresa goto Step 10. |
|
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
|
|
|
|
if ((iMinSAD <= threshA) || |
|
|
(MVequal(*currMV, prevMB->i_mvs[0]) && |
|
|
((int32_t) iMinSAD < prevMB->i_sad16))) { |
|
|
|
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfastInt16_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
|
|
|
/************ (Diamond Search) **************/ |
|
|
/* |
|
|
Step 7: Perform Diamond search, with either the small or large diamond. |
|
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES16) |
|
|
MainSearchPtr = Square16_MainSearch; |
|
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamond16_MainSearch; |
|
|
else |
|
|
MainSearchPtr = Diamond16_MainSearch; |
|
|
|
|
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
|
|
|
|
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, |
|
|
currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, |
|
|
min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_EXTSEARCH16) { |
|
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
|
|
|
|
if (!(MVequal(pmv[0], backupMV))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, |
|
|
min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
|
|
|
if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, |
|
|
iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, |
|
|
max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
/* |
|
|
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
|
|
*/ |
|
|
|
|
|
PMVfastInt16_Terminate_with_Refine: |
|
|
|
|
|
pMB->i_mvs[0] = pMB->i_mvs[1] = pMB->i_mvs[2] = pMB->i_mvs[3] = pMB->i_mv16 = *currMV; |
|
|
pMB->i_sad8[0] = pMB->i_sad8[1] = pMB->i_sad8[2] = pMB->i_sad8[3] = pMB->i_sad16 = iMinSAD; |
|
|
|
|
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
|
|
iMinSAD = |
|
|
Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
|
|
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
|
|
iFcode, iQuant, iEdgedWidth); |
|
|
|
|
|
pmv[0] = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); // get _REAL_ prediction (halfpel possible) |
|
|
|
|
|
PMVfastInt16_Terminate_without_Refine: |
|
|
currPMV->x = currMV->x - center_x; |
|
|
currPMV->y = currMV->y - center_y; |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
/* *********************************************************** |
|
|
bvop motion estimation |
|
|
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
|
|
***************************************************************/ |
|
|
|
|
|
|
|
|
#define DIRECT_PENALTY 0 |
|
|
#define DIRECT_UPPERLIMIT 256 // never use direct mode if SAD is larger than this |
|
|
|
|
|
void |
|
|
MotionEstimationBVOP(MBParam * const pParam, |
|
|
FRAMEINFO * const frame, |
|
|
const int32_t time_bp, |
|
|
const int32_t time_pp, |
|
|
// forward (past) reference |
|
|
const MACROBLOCK * const f_mbs, |
|
|
const IMAGE * const f_ref, |
|
|
const IMAGE * const f_refH, |
|
|
const IMAGE * const f_refV, |
|
|
const IMAGE * const f_refHV, |
|
|
// backward (future) reference |
|
|
const MACROBLOCK * const b_mbs, |
|
|
const IMAGE * const b_ref, |
|
|
const IMAGE * const b_refH, |
|
|
const IMAGE * const b_refV, |
|
|
const IMAGE * const b_refHV) |
|
|
{ |
|
|
const int mb_width = pParam->mb_width; |
|
|
const int mb_height = pParam->mb_height; |
|
|
const int edged_width = pParam->edged_width; |
|
|
|
|
|
int i, j, k; |
|
|
|
|
|
static const VECTOR zeroMV={0,0}; |
|
|
|
|
|
int f_sad16; /* forward (as usual) search */ |
|
|
int b_sad16; /* backward (only in b-frames) search */ |
|
|
int i_sad16; /* interpolated (both direction, b-frames only) */ |
|
|
int d_sad16; /* direct mode (assume linear motion) */ |
|
|
int dnv_sad16; /* direct mode (assume linear motion) without correction vector */ |
|
|
|
|
|
int best_sad; |
|
|
|
|
|
VECTOR f_predMV, b_predMV; /* there is no direct prediction */ |
|
|
VECTOR pmv_dontcare; |
|
|
|
|
|
int f_count=0; |
|
|
int b_count=0; |
|
|
int i_count=0; |
|
|
int d_count=0; |
|
|
int dnv_count=0; |
|
|
int s_count=0; |
|
|
const int64_t TRB = (int32_t)time_pp - (int32_t)time_bp; |
|
|
const int64_t TRD = (int32_t)time_pp; |
|
|
|
|
|
// fprintf(stderr,"TRB = %lld TRD = %lld time_bp =%d time_pp =%d\n\n",TRB,TRD,time_bp,time_pp); |
|
|
// note: i==horizontal, j==vertical |
|
|
for (j = 0; j < mb_height; j++) { |
|
|
|
|
|
f_predMV = zeroMV; /* prediction is reset at left boundary */ |
|
|
b_predMV = zeroMV; |
|
|
|
|
|
for (i = 0; i < mb_width; i++) { |
|
|
MACROBLOCK *mb = &frame->mbs[i + j * mb_width]; |
|
|
const MACROBLOCK *f_mb = &f_mbs[i + j * mb_width]; |
|
|
const MACROBLOCK *b_mb = &b_mbs[i + j * mb_width]; |
|
|
|
|
|
VECTOR directMV; |
|
|
VECTOR deltaMV=zeroMV; |
|
|
|
|
|
/* special case, if collocated block is SKIPed: encoding is forward(0,0) */ |
|
|
|
|
|
if (b_mb->mode == MODE_INTER && b_mb->cbp == 0 && |
|
|
b_mb->mvs[0].x == 0 && b_mb->mvs[0].y == 0) { |
|
|
mb->mode = MODE_NOT_CODED; |
|
|
mb->mvs[0].x = 0; |
|
|
mb->mvs[0].y = 0; |
|
|
mb->b_mvs[0].x = 0; |
|
|
mb->b_mvs[0].y = 0; |
|
|
continue; |
|
|
} |
|
|
|
|
|
dnv_sad16 = DIRECT_PENALTY; |
|
|
|
|
|
if (b_mb->mode == MODE_INTER4V) |
|
|
{ |
|
|
|
|
|
/* same method of scaling as in decoder.c, so we copy from there */ |
|
|
for (k = 0; k < 4; k++) { |
|
|
|
|
|
directMV = b_mb->mvs[k]; |
|
|
|
|
|
mb->mvs[k].x = (int32_t) ((TRB * directMV.x) / TRD + deltaMV.x); |
|
|
mb->b_mvs[k].x = (int32_t) ((deltaMV.x == 0) |
|
|
? ((TRB - TRD) * directMV.x) / TRD |
|
|
: mb->mvs[k].x - directMV.x); |
|
|
mb->mvs[k].y = (int32_t) ((TRB * directMV.y) / TRD + deltaMV.y); |
|
|
mb->b_mvs[k].y = (int32_t) ((deltaMV.y == 0) |
|
|
? ((TRB - TRD) * directMV.y) / TRD |
|
|
: mb->mvs[k].y - directMV.y); |
|
|
|
|
|
dnv_sad16 += |
|
|
sad8bi(frame->image.y + 2*(i+(k&1))*8 + 2*(j+(k>>1))*8*edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
2*(i+(k&1)), 2*(j+(k>>1)), 8, &mb->mvs[k], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
2*(i+(k&1)), 2*(j+(k>>1)), 8, &mb->b_mvs[k], edged_width), |
|
|
edged_width); |
|
|
} |
|
|
} |
|
|
else |
|
|
{ |
|
|
directMV = b_mb->mvs[0]; |
|
|
|
|
|
mb->mvs[0].x = (int32_t) ((TRB * directMV.x) / TRD + deltaMV.x); |
|
|
mb->b_mvs[0].x = (int32_t) ((deltaMV.x == 0) |
|
|
? ((TRB - TRD) * directMV.x) / TRD |
|
|
: mb->mvs[0].x - directMV.x); |
|
|
mb->mvs[0].y = (int32_t) ((TRB * directMV.y) / TRD + deltaMV.y); |
|
|
mb->b_mvs[0].y = (int32_t) ((deltaMV.y == 0) |
|
|
? ((TRB - TRD) * directMV.y) / TRD |
|
|
: mb->mvs[0].y - directMV.y); |
|
|
|
|
|
dnv_sad16 = DIRECT_PENALTY + |
|
|
sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, 16, &mb->mvs[0], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, &mb->b_mvs[0], edged_width), |
|
|
edged_width); |
|
|
|
|
|
|
|
|
} |
|
|
|
|
|
// forward search |
|
|
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
&frame->image, i, j, |
|
|
mb->mvs[0].x, mb->mvs[0].y, /* start point f_directMV */ |
|
|
f_predMV.x, f_predMV.y, /* center is f-prediction */ |
|
|
frame->motion_flags & (~(PMV_EARLYSTOP16|PMV_QUICKSTOP16)), |
|
|
frame->quant, frame->fcode, pParam, |
|
|
f_mbs, f_mbs, |
|
|
&mb->mvs[0], &pmv_dontcare); |
|
|
|
|
|
|
|
|
// backward search |
|
|
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
&frame->image, i, j, |
|
|
mb->b_mvs[0].x, mb->b_mvs[0].y, /* start point b_directMV */ |
|
|
b_predMV.x, b_predMV.y, /* center is b-prediction */ |
|
|
frame->motion_flags & (~(PMV_EARLYSTOP16|PMV_QUICKSTOP16)), |
|
|
frame->quant, frame->bcode, pParam, |
|
|
b_mbs, b_mbs, |
|
|
&mb->b_mvs[0], &pmv_dontcare); |
|
|
|
|
|
i_sad16 = |
|
|
sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, 16, &mb->mvs[0], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, &mb->b_mvs[0], edged_width), |
|
|
edged_width); |
|
|
i_sad16 += calc_delta_16(mb->mvs[0].x-f_predMV.x, mb->mvs[0].y-f_predMV.y, |
|
|
frame->fcode, frame->quant); |
|
|
i_sad16 += calc_delta_16(mb->b_mvs[0].y-b_predMV.y, mb->b_mvs[0].y-b_predMV.y, |
|
|
frame->bcode, frame->quant); |
|
|
|
|
|
// TODO: direct search |
|
|
// predictor + delta vector in range [-32,32] (fcode=1) |
|
|
|
|
|
// i_sad16 = 65535; |
|
|
// f_sad16 = 65535; |
|
|
// b_sad16 = 65535; |
|
|
|
|
|
if (f_sad16 < b_sad16) { |
|
|
best_sad = f_sad16; |
|
|
mb->mode = MODE_FORWARD; |
|
|
} else { |
|
|
best_sad = b_sad16; |
|
|
mb->mode = MODE_BACKWARD; |
|
|
} |
|
|
|
|
|
if (i_sad16 < best_sad) { |
|
|
best_sad = i_sad16; |
|
|
mb->mode = MODE_INTERPOLATE; |
|
|
} |
|
|
|
|
|
if (dnv_sad16 < best_sad) { |
|
|
|
|
|
if (dnv_sad16 > DIRECT_UPPERLIMIT) |
|
|
{ |
|
|
/* if SAD value is too large, try same vector with MODE_INTERPOLATE |
|
|
instead (interpolate has residue encoding, direct mode without MV |
|
|
doesn't) |
|
|
|
|
|
This has to be replaced later by "real" direct mode, including delta |
|
|
vector and (if needed) residue encoding |
|
|
|
|
|
*/ |
|
|
|
|
|
directMV = b_mb->mvs[0]; |
|
|
|
|
|
mb->mvs[0].x = (int32_t) ((TRB * directMV.x) / TRD + deltaMV.x); |
|
|
mb->b_mvs[0].x = (int32_t) ((deltaMV.x == 0) |
|
|
? ((TRB - TRD) * directMV.x) / TRD |
|
|
: mb->mvs[0].x - directMV.x); |
|
|
mb->mvs[0].y = (int32_t) ((TRB * directMV.y) / TRD + deltaMV.y); |
|
|
mb->b_mvs[0].y = (int32_t) ((deltaMV.y == 0) |
|
|
? ((TRB - TRD) * directMV.y) / TRD |
|
|
: mb->mvs[0].y - directMV.y); |
|
|
|
|
|
dnv_sad16 = |
|
|
sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, 16, &mb->mvs[0], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, &mb->b_mvs[0], edged_width), |
|
|
edged_width); |
|
|
dnv_sad16 += calc_delta_16(mb->mvs[0].x-f_predMV.x, mb->mvs[0].y-f_predMV.y, |
|
|
frame->fcode, frame->quant); |
|
|
dnv_sad16 += calc_delta_16(mb->b_mvs[0].y-b_predMV.y, mb->b_mvs[0].y-b_predMV.y, |
|
|
frame->bcode, frame->quant); |
|
|
|
|
|
if (dnv_sad16 < best_sad) |
|
|
{ |
|
|
best_sad = dnv_sad16; |
|
|
mb->mode = MODE_INTERPOLATE; |
|
|
|
|
|
/* fprintf(stderr,"f_sad16 = %d, b_sad16 = %d, i_sad16 = %d, dnv_sad16 = %d\n", |
|
|
f_sad16,b_sad16,i_sad16,dnv_sad16); |
|
|
*/ } |
|
|
} |
|
|
else |
|
|
{ |
|
|
best_sad = dnv_sad16; |
|
|
mb->mode = MODE_DIRECT_NONE_MV; |
|
|
} |
|
|
} |
|
|
|
|
|
switch (mb->mode) |
|
|
{ |
|
|
case MODE_FORWARD: |
|
|
f_count++; |
|
|
f_predMV = mb->mvs[0]; |
|
|
break; |
|
|
case MODE_BACKWARD: |
|
|
b_count++; |
|
|
b_predMV = mb->b_mvs[0]; |
|
|
|
|
|
break; |
|
|
case MODE_INTERPOLATE: |
|
|
i_count++; |
|
|
f_predMV = mb->mvs[0]; |
|
|
b_predMV = mb->b_mvs[0]; |
|
|
break; |
|
|
case MODE_DIRECT: |
|
|
d_count++; break; |
|
|
case MODE_DIRECT_NONE_MV: |
|
|
dnv_count++; break; |
|
|
default: |
|
|
s_count++; break; |
|
|
} |
|
|
|
|
|
} |
|
|
} |
|
|
|
|
|
#ifdef _DEBUG_BFRAME_STAT |
|
|
fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D0: %04d D: %04d S: %04d\n", |
|
|
f_count,b_count,i_count,dnv_count,d_count,s_count); |
|
|
#endif |
|
|
|
|
|
} |
|
Parent Directory
| 
Revision Log
| 
Patch