--- trunk/xvidcore/src/motion/motion_est.c 2002/07/25 00:43:19 341 +++ trunk/xvidcore/src/motion/motion_est.c 2002/09/22 20:34:05 524 @@ -1,69 +1,36 @@ -/************************************************************************** +/***************************************************************************** * - * XVID MPEG-4 VIDEO CODEC - * motion estimation + * XVID MPEG-4 VIDEO CODEC + * - Motion Estimation module - * - * This program is an implementation of a part of one or more MPEG-4 - * Video tools as specified in ISO/IEC 14496-2 standard. Those intending - * to use this software module in hardware or software products are - * advised that its use may infringe existing patents or copyrights, and - * any such use would be at such party's own risk. The original - * developer of this software module and his/her company, and subsequent - * editors and their companies, will have no liability for use of this - * software or modifications or derivatives thereof. + * Copyright(C) 2002 Christoph Lampert + * Copyright(C) 2002 Michael Militzer + * Copyright(C) 2002 Edouard Gomez * - * This program is free software; you can redistribute it and/or modify - * it under the terms of the GNU General Public License as published by - * the Free Software Foundation; either version 2 of the License, or - * (at your option) any later version. + * This program is an implementation of a part of one or more MPEG-4 + * Video tools as specified in ISO/IEC 14496-2 standard. Those intending + * to use this software module in hardware or software products are + * advised that its use may infringe existing patents or copyrights, and + * any such use would be at such party's own risk. The original + * developer of this software module and his/her company, and subsequent + * editors and their companies, will have no liability for use of this + * software or modifications or derivatives thereof. * - * This program is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2 of the License, or + * (at your option) any later version. * - * You should have received a copy of the GNU General Public License - * along with this program; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. * - *************************************************************************/ - -/************************************************************************** - * - * Modifications: - * - * 01.05.2002 updated MotionEstimationBVOP - * 25.04.2002 partial prevMB conversion - * 22.04.2002 remove some compile warning by chenm001 - * 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. + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA * - * Michael Militzer - * - **************************************************************************/ + *************************************************************************/ #include #include @@ -174,8 +141,8 @@ static const VECTOR zeroMV = { 0, 0 }; VECTOR predMV; - int32_t x, y; - int32_t iIntra = 0; + uint32_t x, y; + uint32_t iIntra = 0; VECTOR pmv; if (sadInit) @@ -389,8 +356,8 @@ { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ } +#if 0 /* too slow and not fully functional at the moment */ -/* int32_t ZeroSearch16( const uint8_t * const pRef, const uint8_t * const pRefH, @@ -429,7 +396,7 @@ return iSAD; } -*/ +#endif /* 0 */ int32_t Diamond16_MainSearch(const uint8_t * const pRef, @@ -439,12 +406,12 @@ const uint8_t * const cur, const int x, const int y, - const int start_x, - const int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, + const int start_x, + const int start_y, + int iMinSAD, + VECTOR * const currMV, + const int center_x, + const int center_y, const int32_t min_dx, const int32_t max_dx, const int32_t min_dy, @@ -458,6 +425,7 @@ /* Do a diamond search around given starting point, return SAD of best */ int32_t iDirection = 0; + int32_t iDirectionBackup; int32_t iSAD; VECTOR backupMV; @@ -471,23 +439,25 @@ CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); - if (iDirection) + if (iDirection) { while (!iFound) { iFound = 1; backupMV = *currMV; + iDirectionBackup = iDirection; - if (iDirection != 2) + if (iDirectionBackup != 2) CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, 1); - if (iDirection != 1) + if (iDirectionBackup != 1) CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, 2); - if (iDirection != 4) + if (iDirectionBackup != 4) CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, 3); - if (iDirection != 3) + if (iDirectionBackup != 3) CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, 4); + } } else { currMV->x = start_x; currMV->y = start_y; @@ -503,12 +473,12 @@ const uint8_t * const cur, const int x, const int y, - const int start_x, - const int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, + const int start_x, + const int start_y, + int iMinSAD, + VECTOR * const currMV, + const int center_x, + const int center_y, const int32_t min_dx, const int32_t max_dx, const int32_t min_dy, @@ -551,7 +521,7 @@ backupMV.y + iDiamondSize, 8); - if (iDirection) + if (iDirection) { while (!iFound) { iFound = 1; backupMV = *currMV; @@ -560,62 +530,62 @@ case 1: CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y - iDiamondSize, 7); break; case 2: - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y + iDiamondSize, 8); break; case 3: - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, 4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y + iDiamondSize, 8); break; case 4: - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, 3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y + iDiamondSize, 6); break; case 5: - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, 3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y - iDiamondSize, 7); break; case 6: - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, 3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y + iDiamondSize, 8); break; @@ -623,48 +593,49 @@ case 7: CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, 4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y + iDiamondSize, 8); break; case 8: - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, 4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y + iDiamondSize, 8); break; default: - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, 3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, 4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y + iDiamondSize, 8); break; } + } } else { currMV->x = start_x; currMV->y = start_y; @@ -681,12 +652,12 @@ const uint8_t * const cur, const int x, const int y, - const int start_x, - const int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, + const int start_x, + const int start_y, + int iMinSAD, + VECTOR * const currMV, + const int center_x, + const int center_y, const int32_t min_dx, const int32_t max_dx, const int32_t min_dy, @@ -719,12 +690,12 @@ const uint8_t * const cur, const int x, const int y, - int start_x, - int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, + const int start_xi, + const int start_yi, + int iMinSAD, + VECTOR * const currMV, + const int center_x, + const int center_y, const int32_t min_dx, const int32_t max_dx, const int32_t min_dy, @@ -737,6 +708,7 @@ { int32_t iSAD; + int start_x = start_xi, start_y = start_yi; /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ @@ -867,164 +839,9 @@ } while (1); //forever } - 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, - - 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; - } return iMinSAD; } -#endif - int32_t AdvDiamond8_MainSearch(const uint8_t * const pRef, @@ -1034,8 +851,8 @@ const uint8_t * const cur, const int x, const int y, - int start_x, - int start_y, + const int start_xi, + const int start_yi, int iMinSAD, VECTOR * const currMV, const int center_x, @@ -1052,6 +869,7 @@ { int32_t iSAD; + int start_x = start_xi, start_y = start_yi; /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ @@ -1274,9 +1092,9 @@ const IMAGE * const pCur, const int x, const int y, - const int start_x, - const int start_y, - const int center_x, + const int start_x, /* start is searched first, so it should contain the most */ + const int start_y, /* likely motion vector for this block */ + const int center_x, /* center is from where length of MVs is measured */ const int center_y, const uint32_t MotionFlags, const uint32_t iQuant, @@ -1331,7 +1149,6 @@ } /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - //bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); if ((x == 0) && (y == 0)) { @@ -1389,7 +1206,7 @@ if ((iMinSAD < 256) || ((MVequal(*currMV, prevMB->mvs[0])) && ((int32_t) iMinSAD < prevMB->sad16))) { - if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode + if (iMinSAD < (int)(2 * iQuant)) // high chances for SKIP-mode { if (!MVzero(*currMV)) { iMinSAD += MV16_00_BIAS; @@ -1594,12 +1411,12 @@ const uint8_t * const cur, const int x, const int y, - int32_t start_x, - int32_t start_y, + const int32_t start_x, + const int32_t start_y, int32_t iMinSAD, VECTOR * const currMV, - const int center_x, - const int center_y, + const int center_x, + const int center_y, const int32_t min_dx, const int32_t max_dx, const int32_t min_dy, @@ -1613,6 +1430,7 @@ /* Do a diamond search around given starting point, return SAD of best */ int32_t iDirection = 0; + int32_t iDirectionBackup; int32_t iSAD; VECTOR backupMV; @@ -1626,23 +1444,25 @@ CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); - if (iDirection) + if (iDirection) { while (!iFound) { iFound = 1; backupMV = *currMV; // since iDirection!=0, this is well defined! + iDirectionBackup = iDirection; - if (iDirection != 2) + if (iDirectionBackup != 2) CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, 1); - if (iDirection != 1) + if (iDirectionBackup != 1) CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, 2); - if (iDirection != 4) + if (iDirectionBackup != 4) CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, 3); - if (iDirection != 3) + if (iDirectionBackup != 3) CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, 4); + } } else { currMV->x = start_x; currMV->y = start_y; @@ -1650,97 +1470,282 @@ return iMinSAD; } -int32_t -Halfpel8_Refine_c(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - VECTOR * const currMV, - int32_t iMinSAD, - const int center_x, - const int 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 iFcode, - const int32_t iQuant, - const int32_t iEdgedWidth) -{ -/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ - - int32_t iSAD; - VECTOR backupMV = *currMV; - - 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); - - return iMinSAD; -} -#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) int32_t -PMVfastSearch8(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) +Square8_MainSearch(const uint8_t * const pRef, + const uint8_t * const pRefH, + const uint8_t * const pRefV, + const uint8_t * const pRefHV, + const uint8_t * const cur, + const int x, + const int y, + const int32_t start_x, + const int32_t start_y, + int32_t iMinSAD, + VECTOR * const currMV, + const int center_x, + const int 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 iFcode, + const int32_t iQuant, + int iFound) { - 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; +/* Do a square search around given starting point, return SAD of best */ - const uint8_t *cur = pCur->y + x * 8 + y * 8 * iEdgedWidth; + int32_t iDirection = 0; + int32_t iSAD; + VECTOR backupMV; - int32_t iDiamondSize; + backupMV.x = start_x; + backupMV.y = start_y; - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; +/* It's one search with full square pattern, and new parts for all following diamonds */ - VECTOR pmv[4]; - int32_t psad[4]; - VECTOR newMV; - VECTOR backupMV; - VECTOR startMV; +/* new direction are extra, so 1-4 is normal diamond + 537 + 1*2 + 648 +*/ -// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; - const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; + CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, 1); + CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, 2); + CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); + CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); - int32_t threshA, threshB; - int32_t iFound, bPredEq; - int32_t iMinSAD, iSAD; + CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); - int32_t iSubBlock = (y & 1) + (y & 1) + (x & 1); - MainSearch8FuncPtr MainSearchPtr; + if (iDirection) { + while (!iFound) { + iFound = 1; + backupMV = *currMV; + + switch (iDirection) { + case 1: + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y, 1); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + break; + case 2: + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, + 2); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + + case 3: + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, + 4); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + + case 4: + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, + 3); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + break; + + case 5: + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, + 1); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, + 3); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + break; + + case 6: + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, + 2); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, + 3); + + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + + break; + + case 7: + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y, 1); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, + 4); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + + case 8: + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, + 2); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, + 4); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + default: + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, + 1); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, + 2); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, + 3); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, + 4); + + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + } + } + } else { + currMV->x = start_x; + currMV->y = start_y; + } + return iMinSAD; +} + + + + + +int32_t +Halfpel8_Refine_c(const uint8_t * const pRef, + const uint8_t * const pRefH, + const uint8_t * const pRefV, + const uint8_t * const pRefHV, + const uint8_t * const cur, + const int x, + const int y, + VECTOR * const currMV, + int32_t iMinSAD, + const int center_x, + const int 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 iFcode, + const int32_t iQuant, + const int32_t iEdgedWidth) +{ +/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ + + int32_t iSAD; + VECTOR backupMV = *currMV; + + 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); + + return iMinSAD; +} + + +#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) + +int32_t +PMVfastSearch8(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 * 8 + y * 8 * iEdgedWidth; + + int32_t iDiamondSize; + + int32_t min_dx; + int32_t max_dx; + int32_t min_dy; + int32_t max_dy; + + VECTOR pmv[4]; + int32_t psad[4]; + VECTOR newMV; + VECTOR backupMV; + VECTOR startMV; + +// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; + const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; + + int32_t threshA, threshB; + int32_t iFound, bPredEq; + int32_t iMinSAD, iSAD; + + int32_t iSubBlock = (y & 1) + (y & 1) + (x & 1); + + MainSearch8FuncPtr MainSearchPtr; /* Init variables */ startMV.x = start_x; @@ -1758,7 +1763,6 @@ } /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ - //bPredEq = get_pmvdata(pMBs, (x >> 1), (y >> 1), iWcount, iSubBlock, pmv, psad); bPredEq = get_pmvdata2(pMBs, iWcount, 0, (x >> 1), (y >> 1), iSubBlock, pmv, psad); if ((x == 0) && (y == 0)) { @@ -1788,9 +1792,9 @@ // Prepare for main loop -// if (MotionFlags & PMV_USESQUARES8) -// MainSearchPtr = Square8_MainSearch; -// else + if (MotionFlags & PMV_USESQUARES8) + MainSearchPtr = Square8_MainSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamond8_MainSearch; @@ -2073,7 +2077,6 @@ max_dy = EVEN(max_dy); } /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - //bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); /* Step 4: Calculate SAD around the Median prediction. @@ -2356,7 +2359,6 @@ max_dy = EVEN(max_dy); } /* 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); bPredEq = get_pmvdata2(pMBs, iWcount, 0, x >> 1, y >> 1, iSubBlock, pmv, psad); @@ -2468,9 +2470,9 @@ // there is no EPZS^2 for inter4v at the moment -// if (MotionFlags & PMV_USESQUARES8) -// MainSearchPtr = Square8_MainSearch; -// else + if (MotionFlags & PMV_USESQUARES8) + MainSearchPtr = Square8_MainSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamond8_MainSearch; @@ -2534,563 +2536,3 @@ currPMV->y = currMV->y - center_y; return iMinSAD; } - - - -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 MinSADx > 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 MinSADi_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 best_sad; - - VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ - VECTOR pmv_dontcare; - - int f_count=0; - int b_count=0; - int i_count=0; - int d_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]; - - mb->deltamv=zeroMV; - -/* special case, if collocated block is SKIPed: encoding is forward(0,0) */ - -#ifndef _DISABLE_SKIP - 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; - } -#endif - - d_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++) { - - mb->directmv[k] = b_mb->mvs[k]; - - mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); - mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) - ? ((TRB - TRD) * mb->directmv[k].x) / TRD - : mb->mvs[k].x - mb->directmv[k].x); - - mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); - mb->b_mvs[k].y = (int32_t) ((mb->directmv[k].y == 0) - ? ((TRB - TRD) * mb->directmv[k].y) / TRD - : mb->mvs[k].y - mb->directmv[k].y); - - d_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 - { - mb->directmv[3] = mb->directmv[2] = mb->directmv[1] = - mb->directmv[0] = b_mb->mvs[0]; - - mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); - mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) - ? ((TRB - TRD) * mb->directmv[0].x) / TRD - : mb->mvs[0].x - mb->directmv[0].x); - - mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); - mb->b_mvs[0].y = (int32_t) ((mb->directmv[0].y == 0) - ? ((TRB - TRD) * mb->directmv[0].y) / TRD - : mb->mvs[0].y - mb->directmv[0].y); - - d_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); - - } - d_sad16 += calc_delta_16(mb->deltamv.x, mb->deltamv.y, 1, frame->quant); - - // 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, - 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, - 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].x-b_predMV.x, 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; -// d_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 (d_sad16 < best_sad) { - - 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++) { - - mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); - mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) - ? ((TRB - TRD) * mb->directmv[k].x) / TRD - : mb->mvs[k].x - mb->directmv[k].x); - - mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); - mb->b_mvs[k].y = (int32_t) ((mb->directmv[k].y == 0) - ? ((TRB - TRD) * mb->directmv[k].y) / TRD - : mb->mvs[k].y - mb->directmv[k].y); - } - } - else - { - mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); - - mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) - ? ((TRB - TRD) * mb->directmv[0].x) / TRD - : mb->mvs[0].x - mb->directmv[0].x); - - mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); - - mb->b_mvs[0].y = (int32_t) ((mb->directmv[0].y == 0) - ? ((TRB - TRD) * mb->directmv[0].y) / TRD - : mb->mvs[0].y - mb->directmv[0].y); - - mb->mvs[3] = mb->mvs[2] = mb->mvs[1] = mb->mvs[0]; - mb->b_mvs[3] = mb->b_mvs[2] = mb->b_mvs[1] = mb->b_mvs[0]; - } - - best_sad = d_sad16; - mb->mode = MODE_DIRECT; - mb->mode = MODE_INTERPOLATE; // direct mode still broken :-( - } - - 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; - default: - s_count++; // ??? - break; - } - - } - } - -#ifdef _DEBUG_BFRAME_STAT - fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D: %04d S: %04d\n", - f_count,b_count,i_count,d_count,s_count); -#endif - -}