--- trunk/xvidcore/src/motion/motion_est.c 2002/06/12 20:38:41 195 +++ trunk/xvidcore/src/motion/motion_est.c 2002/09/07 13:41:13 444 @@ -1,39 +1,36 @@ -/************************************************************************** +/***************************************************************************** * - * Modifications: + * XVID MPEG-4 VIDEO CODEC + * - Motion Estimation module - * - * 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. + * Copyright(C) 2002 Christoph Lampert + * Copyright(C) 2002 Michael Militzer + * Copyright(C) 2002 Edouard Gomez * - * Michael Militzer + * 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 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 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. + * + * 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 + * + *************************************************************************/ #include #include @@ -47,156 +44,7 @@ #include "motion.h" #include "sad.h" -// very large value -#define MV_MAX_ERROR (4096 * 256) - -// stop search if sdelta < THRESHOLD -#define MV16_THRESHOLD 192 -#define MV8_THRESHOLD 56 - -#define NEIGH_MOVE_THRESH 0 -// how much a block's MV must differ from his neighbour -// to be search for INTER4V. The more, the faster... - -/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ -/* nb = vop pixels * 2^(bpp-8) */ -#define MV16_00_BIAS (128+1) -#define MV8_00_BIAS (0) - -/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ -#define MV16_INTER_BIAS 512 - -/* Parameters which control inter/inter4v decision */ -#define IMV16X16 5 - -/* vector map (vlc delta size) smoother parameters */ -#define NEIGH_TEND_16X16 2 -#define NEIGH_TEND_8X8 2 - -// fast ((A)/2)*2 -#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) - -#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) -#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) - -int32_t PMVfastSearch16(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 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); - -int32_t EPZSSearch16(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 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); - - -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 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); - -int32_t EPZSSearch8(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 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); - - -typedef int32_t(MainSearch16Func) (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, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - 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); - -typedef MainSearch16Func *MainSearch16FuncPtr; - - -typedef int32_t(MainSearch8Func) (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, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - 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); -typedef MainSearch8Func *MainSearch8FuncPtr; static int32_t lambda_vec16[32] = /* rounded values for lambda param for weight of motion bits as in modified H.26L */ { 0, (int) (1.00235 + 0.5), (int) (1.15582 + 0.5), (int) (1.31976 + 0.5), @@ -273,22 +121,7 @@ return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); } - - - - - -#ifndef SEARCH16 -#define SEARCH16 PMVfastSearch16 -//#define SEARCH16 FullSearch16 -//#define SEARCH16 EPZSSearch16 -#endif - -#ifndef SEARCH8 -#define SEARCH8 PMVfastSearch8 -//#define SEARCH8 EPZSSearch8 -#endif - + bool MotionEstimation(MBParam * const pParam, FRAMEINFO * const current, @@ -305,7 +138,8 @@ const IMAGE *const pCurrent = ¤t->image; const IMAGE *const pRef = &reference->image; - const VECTOR zeroMV = { 0, 0 }; + static const VECTOR zeroMV = { 0, 0 }; + VECTOR predMV; int32_t x, y; int32_t iIntra = 0; @@ -314,13 +148,20 @@ if (sadInit) (*sadInit) (); - for (y = 0; y < iHcount; y++) - for (x = 0; x < iWcount; x++) { + for (y = 0; y < iHcount; y++) { + for (x = 0; x < iWcount; x ++) { + MACROBLOCK *const pMB = &pMBs[x + y * iWcount]; + if (pMB->mode == MODE_NOT_CODED) + continue; + + predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); + pMB->sad16 = - SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, - y, current->motion_flags, current->quant, + SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, + x, y, predMV.x, predMV.y, predMV.x, predMV.y, + current->motion_flags, current->quant, current->fcode, pParam, pMBs, prevMBs, &pMB->mv16, &pMB->pmvs[0]); @@ -352,43 +193,52 @@ pMB->dquant == NO_CHANGE)) { int32_t sad8 = IMV16X16 * current->quant; - if (sad8 < pMB->sad16) - + if (sad8 < pMB->sad16) { sad8 += pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, - pCurrent, 2 * x, 2 * y, pMB->mv16.x, - pMB->mv16.y, current->motion_flags, + pCurrent, 2 * x, 2 * y, + pMB->mv16.x, pMB->mv16.y, predMV.x, predMV.y, + current->motion_flags, current->quant, current->fcode, pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); + } + if (sad8 < pMB->sad16) { - if (sad8 < pMB->sad16) + predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 1); sad8 += pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, - pCurrent, 2 * x + 1, 2 * y, pMB->mv16.x, - pMB->mv16.y, current->motion_flags, + pCurrent, 2 * x + 1, 2 * y, + pMB->mv16.x, pMB->mv16.y, predMV.x, predMV.y, + current->motion_flags, current->quant, current->fcode, pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); - - if (sad8 < pMB->sad16) + } + if (sad8 < pMB->sad16) { + predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 2); sad8 += pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, - pCurrent, 2 * x, 2 * y + 1, pMB->mv16.x, - pMB->mv16.y, current->motion_flags, + pCurrent, 2 * x, 2 * y + 1, + pMB->mv16.x, pMB->mv16.y, predMV.x, predMV.y, + current->motion_flags, current->quant, current->fcode, pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); - - if (sad8 < pMB->sad16) + } + if (sad8 < pMB->sad16) { + predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 3); sad8 += pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, 2 * x + 1, 2 * y + 1, - pMB->mv16.x, pMB->mv16.y, - current->motion_flags, current->quant, - current->fcode, pParam, pMBs, prevMBs, - &pMB->mvs[3], &pMB->pmvs[3]); - + pMB->mv16.x, pMB->mv16.y, predMV.x, predMV.y, + current->motion_flags, + current->quant, current->fcode, pParam, + pMBs, prevMBs, + &pMB->mvs[3], + &pMB->pmvs[3]); + } + /* decide: MODE_INTER or MODE_INTER4V mpeg4: if (sad8 < pMB->sad16 - nb/2+1) use_inter4v */ @@ -409,24 +259,26 @@ pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = pMB->sad16; + } + } - } return 0; } + #define CHECK_MV16_ZERO {\ if ( (0 <= max_dx) && (0 >= min_dx) \ && (0 <= max_dy) && (0 >= min_dy) ) \ { \ iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ - iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode, iQuant);\ + iSAD += calc_delta_16(-center_x, -center_y, (uint8_t)iFcode, iQuant);\ if (iSAD < iMinSAD) \ { iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ } #define NOCHECK_MV16_CANDIDATE(X,Y) { \ iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ - iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ + 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); } \ } @@ -436,7 +288,7 @@ && ((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) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ + 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); } } \ } @@ -446,7 +298,7 @@ && ((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) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ + 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); } } \ } @@ -456,7 +308,7 @@ && ((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) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ + 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; } } \ } @@ -464,7 +316,7 @@ #define CHECK_MV8_ZERO {\ iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ - iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode, iQuant);\ + iSAD += calc_delta_8(-center_x, -center_y, (uint8_t)iFcode, iQuant);\ if (iSAD < iMinSAD) \ { iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ } @@ -472,7 +324,7 @@ #define NOCHECK_MV8_CANDIDATE(X,Y) \ { \ iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ + iSAD += calc_delta_8((X)-center_x, (Y)-center_y, (uint8_t)iFcode, iQuant);\ if (iSAD < iMinSAD) \ { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ } @@ -482,7 +334,7 @@ && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ { \ iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ + iSAD += calc_delta_8((X)-center_x, (Y)-center_y, (uint8_t)iFcode, iQuant);\ if (iSAD < iMinSAD) \ { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ } @@ -492,7 +344,7 @@ && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ { \ iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ + iSAD += calc_delta_8((X)-center_x, (Y)-center_y, (uint8_t)iFcode, iQuant);\ if (iSAD < iMinSAD) \ { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ } @@ -502,7 +354,7 @@ && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ { \ iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ + iSAD += calc_delta_8((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; } } \ } @@ -528,9 +380,10 @@ const int32_t iEdgedWidth = pParam->edged_width; const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; int32_t iSAD; - int32_t pred_x,pred_y; + VECTOR pred; + - get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); + pred = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), @@ -540,8 +393,8 @@ currMV->x = 0; currMV->y = 0; - currPMV->x = -pred_x; - currPMV->y = -pred_y; + currPMV->x = -pred.x; + currPMV->y = -pred.y; return iSAD; @@ -556,11 +409,12 @@ const uint8_t * const cur, const int x, const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, + 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, @@ -574,11 +428,12 @@ /* Do a diamond search around given starting point, return SAD of best */ int32_t iDirection = 0; + int32_t iDirectionBackup; int32_t iSAD; VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; + backupMV.x = start_x; + backupMV.y = start_y; /* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ @@ -587,26 +442,28 @@ 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 = startx; - currMV->y = starty; + currMV->x = start_x; + currMV->y = start_y; } return iMinSAD; } @@ -619,11 +476,12 @@ const uint8_t * const cur, const int x, const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, + 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, @@ -640,8 +498,8 @@ int32_t iSAD; VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; + backupMV.x = start_x; + backupMV.y = start_y; /* It's one search with full square pattern, and new parts for all following diamonds */ @@ -666,7 +524,7 @@ backupMV.y + iDiamondSize, 8); - if (iDirection) + if (iDirection) { while (!iFound) { iFound = 1; backupMV = *currMV; @@ -675,62 +533,62 @@ 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; @@ -738,51 +596,52 @@ 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 = startx; - currMV->y = starty; + currMV->x = start_x; + currMV->y = start_y; } return iMinSAD; } @@ -796,11 +655,12 @@ const uint8_t * const cur, const int x, const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, + 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, @@ -815,8 +675,8 @@ int32_t dx, dy; VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; + backupMV.x = start_x; + backupMV.y = start_y; for (dx = min_dx; dx <= max_dx; dx += iDiamondSize) for (dy = min_dy; dy <= max_dy; dy += iDiamondSize) @@ -833,11 +693,12 @@ const uint8_t * const cur, const int x, const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, + int start_x, + 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, @@ -854,26 +715,26 @@ /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ if (iDirection) { - CHECK_MV16_CANDIDATE(startx - iDiamondSize, starty); - CHECK_MV16_CANDIDATE(startx + iDiamondSize, starty); - CHECK_MV16_CANDIDATE(startx, starty - iDiamondSize); - CHECK_MV16_CANDIDATE(startx, starty + iDiamondSize); + CHECK_MV16_CANDIDATE(start_x - iDiamondSize, start_y); + CHECK_MV16_CANDIDATE(start_x + iDiamondSize, start_y); + CHECK_MV16_CANDIDATE(start_x, start_y - iDiamondSize); + CHECK_MV16_CANDIDATE(start_x, start_y + iDiamondSize); } else { int bDirection = 1 + 2 + 4 + 8; do { iDirection = 0; if (bDirection & 1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, starty, 1); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); if (bDirection & 2) - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, starty, 2); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); if (bDirection & 4) - CHECK_MV16_CANDIDATE_DIR(startx, starty - iDiamondSize, 4); + CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); if (bDirection & 8) - CHECK_MV16_CANDIDATE_DIR(startx, starty + iDiamondSize, 8); + CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); /* now we're doing diagonal checks near our candidate */ @@ -881,99 +742,100 @@ { bDirection = iDirection; iDirection = 0; - startx = currMV->x; - starty = currMV->y; + start_x = currMV->x; + start_y = currMV->y; if (bDirection & 3) //our candidate is left or right { - CHECK_MV16_CANDIDATE_DIR(startx, starty + iDiamondSize, 8); - CHECK_MV16_CANDIDATE_DIR(startx, starty - iDiamondSize, 4); + CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); + CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); } else // what remains here is up or down { - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, starty, 2); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, starty, 1); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); } if (iDirection) { bDirection += iDirection; - startx = currMV->x; - starty = currMV->y; + start_x = currMV->x; + start_y = currMV->y; } } else //about to quit, eh? not so fast.... { switch (bDirection) { case 2: - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); break; case 1: - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); + + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); break; case 2 + 4: - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); break; case 4: - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); break; case 8: - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); break; case 1 + 4: - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); break; case 2 + 8: - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); break; case 1 + 8: - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); break; default: //1+2+4+8 == we didn't find anything at all - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); break; } if (!iDirection) break; //ok, the end. really else { bDirection = iDirection; - startx = currMV->x; - starty = currMV->y; + start_x = currMV->x; + start_y = currMV->y; } } } @@ -982,6 +844,252 @@ return iMinSAD; } +#define CHECK_MV16_F_INTERPOL(X,Y) { \ + if ( ((X) <= f_max_dx) && ((X) >= f_min_dx) \ + && ((Y) <= f_max_dy) && ((Y) >= f_min_dy) ) \ + { \ + iSAD = sad16bi( cur, \ + get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ + get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, b_currMV->x, b_currMV->y, iEdgedWidth), \ + iEdgedWidth); \ + iSAD += calc_delta_16((X) - f_center_x, (Y) - f_center_y, (uint8_t)f_iFcode, iQuant);\ + iSAD += calc_delta_16(b_currMV->x - b_center_x, b_currMV->y - b_center_y, (uint8_t)b_iFcode, iQuant);\ + if (iSAD < iMinSAD) \ + { iMinSAD=iSAD; f_currMV->x=(X); f_currMV->y=(Y); } } \ +} + +#define CHECK_MV16_F_INTERPOL_FOUND(X,Y) { \ + if ( ((X) <= f_max_dx) && ((X) >= f_min_dx) \ + && ((Y) <= f_max_dy) && ((Y) >= f_min_dy) ) \ + { \ + iSAD = sad16bi( cur, \ + get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ + get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, b_currMV->x, b_currMV->y, iEdgedWidth), \ + iEdgedWidth); \ + iSAD += calc_delta_16((X) - f_center_x, (Y) - f_center_y, (uint8_t)f_iFcode, iQuant);\ + iSAD += calc_delta_16(b_currMV->x - b_center_x, b_currMV->y - b_center_y, (uint8_t)b_iFcode, iQuant);\ + if (iSAD < iMinSAD) \ + { iMinSAD=iSAD; f_currMV->x=(X); f_currMV->y=(Y); iFound=0;} } \ +} + +#define CHECK_MV16_B_INTERPOL(X,Y) { \ + if ( ((X) <= b_max_dx) && ((X) >= b_min_dx) \ + && ((Y) <= b_max_dy) && ((Y) >= b_min_dy) ) \ + { \ + iSAD = sad16bi( cur, \ + get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, f_currMV->x, f_currMV->y, iEdgedWidth), \ + get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ + iEdgedWidth); \ + iSAD += calc_delta_16(f_currMV->x - f_center_x, f_currMV->y - f_center_y, (uint8_t)f_iFcode, iQuant);\ + iSAD += calc_delta_16((X) - b_center_x, (Y) - b_center_y, (uint8_t)b_iFcode, iQuant);\ + if (iSAD < iMinSAD) \ + { iMinSAD=iSAD; b_currMV->x=(X); b_currMV->y=(Y); } } \ +} + +#define CHECK_MV16_B_INTERPOL_FOUND(X,Y) { \ + if ( ((X) <= b_max_dx) && ((X) >= b_min_dx) \ + && ((Y) <= b_max_dy) && ((Y) >= b_min_dy) ) \ + { \ + iSAD = sad16bi( cur, \ + get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, f_currMV->x, f_currMV->y, iEdgedWidth), \ + get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ + iEdgedWidth); \ + iSAD += calc_delta_16(f_currMV->x - f_center_x, f_currMV->y - f_center_y, (uint8_t)f_iFcode, iQuant);\ + iSAD += calc_delta_16((X) - b_center_x, (Y) - b_center_y, (uint8_t)b_iFcode, iQuant);\ + if (iSAD < iMinSAD) \ + { iMinSAD=iSAD; b_currMV->x=(X); b_currMV->y=(Y); iFound=0;} } \ +} + +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 f_min_dx, + const int32_t f_max_dx, + const int32_t f_min_dy, + const int32_t f_max_dy, + + const int32_t b_min_dx, + const int32_t b_max_dx, + const int32_t b_min_dy, + const int32_t b_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 iSAD; + + VECTOR f_backupMV; + VECTOR b_backupMV; + + f_currMV->x = f_start_x; + f_currMV->y = f_start_y; + b_currMV->x = b_start_x; + b_currMV->y = b_start_y; + + do + { + iFound = 1; + + f_backupMV = *f_currMV; + + CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x - iDiamondSize, f_backupMV.y); + CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x + iDiamondSize, f_backupMV.y); + CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x, f_backupMV.y - iDiamondSize); + CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x, f_backupMV.y + iDiamondSize); + + b_backupMV = *b_currMV; + + CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x - iDiamondSize, b_backupMV.y); + CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x + iDiamondSize, b_backupMV.y); + CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x, b_backupMV.y - iDiamondSize); + CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x, b_backupMV.y + iDiamondSize); + + } while (!iFound); + + return iMinSAD; +} + +/* Sorry, these MACROS really got too large... I'll turn them into function soon! */ + +#define CHECK_MV16_DIRECT_FOUND(X,Y) \ + if ( (X)>=(-32) && (X)<=(31) && ((Y)>=-32) && ((Y)<=31) ) \ + { int k;\ + VECTOR mvs,b_mvs; \ + iSAD = 0;\ + for (k = 0; k < 4; k++) { \ + mvs.x = (int32_t) ((TRB * directmv[k].x) / TRD + (X)); \ + b_mvs.x = (int32_t) (((X) == 0) \ + ? ((TRB - TRD) * directmv[k].x) / TRD \ + : mvs.x - directmv[k].x); \ + \ + mvs.y = (int32_t) ((TRB * directmv[k].y) / TRD + (Y)); \ + b_mvs.y = (int32_t) (((Y) == 0) \ + ? ((TRB - TRD) * directmv[k].y) / TRD \ + : mvs.y - directmv[k].y); \ + \ + if ( (mvs.x <= max_dx) && (mvs.x >= min_dx) \ + && (mvs.y <= max_dy) && (mvs.y >= min_dy) \ + && (b_mvs.x <= max_dx) && (b_mvs.x >= min_dx) \ + && (b_mvs.y <= max_dy) && (b_mvs.y >= min_dy) ) { \ + iSAD += sad8bi( cur + 8*(k&1) + 8*(k>>1)*iEdgedWidth, \ + get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, 2*x+(k&1), 2*y+(k>>1), 8, \ + mvs.x, mvs.y, iEdgedWidth), \ + get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, 2*x+(k&1), 2*y+(k>>1), 8, \ + b_mvs.x, b_mvs.y, iEdgedWidth), \ + iEdgedWidth); \ + } \ + else \ + iSAD = 65535; \ + } \ + iSAD += calc_delta_16((X),(Y), 1, iQuant);\ + if (iSAD < iMinSAD) \ + { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iFound=0; } \ +} + + + +int32_t +Diamond16_DirectMainSearch( + 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 TRB, + const int TRD, + + const int start_x, + const int start_y, + + int iMinSAD, + VECTOR * const currMV, + const VECTOR * const directmv, + + 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 iQuant, + int iFound) +{ +/* Do a diamond search around given starting point, return SAD of best */ + + int32_t iSAD; + + VECTOR backupMV; + + currMV->x = start_x; + currMV->y = start_y; + +/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ + + do + { + iFound = 1; + + backupMV = *currMV; + + CHECK_MV16_DIRECT_FOUND(backupMV.x - iDiamondSize, backupMV.y); + CHECK_MV16_DIRECT_FOUND(backupMV.x + iDiamondSize, backupMV.y); + CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y - iDiamondSize); + CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y + iDiamondSize); + + } while (!iFound); + + return iMinSAD; +} + + int32_t AdvDiamond8_MainSearch(const uint8_t * const pRef, const uint8_t * const pRefH, @@ -990,11 +1098,12 @@ const uint8_t * const cur, const int x, const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, + int start_x, + int start_y, + int iMinSAD, VECTOR * const currMV, - const VECTOR * const pmv, + const int center_x, + const int center_y, const int32_t min_dx, const int32_t max_dx, const int32_t min_dy, @@ -1011,26 +1120,26 @@ /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ if (iDirection) { - CHECK_MV8_CANDIDATE(startx - iDiamondSize, starty); - CHECK_MV8_CANDIDATE(startx + iDiamondSize, starty); - CHECK_MV8_CANDIDATE(startx, starty - iDiamondSize); - CHECK_MV8_CANDIDATE(startx, starty + iDiamondSize); + CHECK_MV8_CANDIDATE(start_x - iDiamondSize, start_y); + CHECK_MV8_CANDIDATE(start_x + iDiamondSize, start_y); + CHECK_MV8_CANDIDATE(start_x, start_y - iDiamondSize); + CHECK_MV8_CANDIDATE(start_x, start_y + iDiamondSize); } else { int bDirection = 1 + 2 + 4 + 8; do { iDirection = 0; if (bDirection & 1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, starty, 1); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); if (bDirection & 2) - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, starty, 2); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); if (bDirection & 4) - CHECK_MV8_CANDIDATE_DIR(startx, starty - iDiamondSize, 4); + CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); if (bDirection & 8) - CHECK_MV8_CANDIDATE_DIR(startx, starty + iDiamondSize, 8); + CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); /* now we're doing diagonal checks near our candidate */ @@ -1038,99 +1147,99 @@ { bDirection = iDirection; iDirection = 0; - startx = currMV->x; - starty = currMV->y; + start_x = currMV->x; + start_y = currMV->y; if (bDirection & 3) //our candidate is left or right { - CHECK_MV8_CANDIDATE_DIR(startx, starty + iDiamondSize, 8); - CHECK_MV8_CANDIDATE_DIR(startx, starty - iDiamondSize, 4); + CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); + CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); } else // what remains here is up or down { - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, starty, 2); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, starty, 1); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); } if (iDirection) { bDirection += iDirection; - startx = currMV->x; - starty = currMV->y; + start_x = currMV->x; + start_y = currMV->y; } } else //about to quit, eh? not so fast.... { switch (bDirection) { case 2: - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); break; case 1: - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); break; case 2 + 4: - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); break; case 4: - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); break; case 8: - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); break; case 1 + 4: - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); break; case 2 + 8: - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); break; case 1 + 8: - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); break; default: //1+2+4+8 == we didn't find anything at all - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y - iDiamondSize, 1 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, + start_y + iDiamondSize, 1 + 8); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y - iDiamondSize, 2 + 4); + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, + start_y + iDiamondSize, 2 + 8); break; } if (!(iDirection)) break; //ok, the end. really else { bDirection = iDirection; - startx = currMV->x; - starty = currMV->y; + start_x = currMV->x; + start_y = currMV->y; } } } @@ -1148,11 +1257,12 @@ const uint8_t * const cur, const int x, const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, + 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, @@ -1167,8 +1277,8 @@ int32_t dx, dy; VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; + backupMV.x = start_x; + backupMV.y = start_y; for (dx = min_dx; dx <= max_dx; dx += iDiamondSize) for (dy = min_dy; dy <= max_dy; dy += iDiamondSize) @@ -1177,7 +1287,7 @@ return iMinSAD; } - +Halfpel8_RefineFuncPtr Halfpel8_Refine; int32_t Halfpel16_Refine(const uint8_t * const pRef, @@ -1189,7 +1299,8 @@ const int y, VECTOR * const currMV, int32_t iMinSAD, - const VECTOR * const pmv, + const int center_x, + const int center_y, const int32_t min_dx, const int32_t max_dx, const int32_t min_dy, @@ -1218,6 +1329,7 @@ #define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) + int32_t PMVfastSearch16(const uint8_t * const pRef, const uint8_t * const pRefH, @@ -1226,6 +1338,10 @@ const IMAGE * const pCur, const int x, const int y, + 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, const uint32_t iFcode, @@ -1259,10 +1375,9 @@ MainSearch16FuncPtr MainSearchPtr; -// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; - static int32_t threshA, threshB; + int32_t threshA, threshB; int32_t bPredEq; int32_t iMinSAD, iSAD; @@ -1280,12 +1395,11 @@ } /* 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)) { threshA = 512; threshB = 1024; - } else { threshA = psad[0]; threshB = threshA + 256; @@ -1306,7 +1420,9 @@ If SAD<=256 goto Step 10. */ - *currMV = pmv[0]; /* current best := prediction */ + currMV->x = start_x; + currMV->y = start_y; + if (!(MotionFlags & PMV_HALFPEL16)) { /* This should NOT be necessary! */ currMV->x = EVEN(currMV->x); currMV->y = EVEN(currMV->y); @@ -1330,12 +1446,12 @@ get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); iMinSAD += - calc_delta_16(currMV->x - pmv[0].x, currMV->y - pmv[0].y, + calc_delta_16(currMV->x - center_x, currMV->y - center_y, (uint8_t) iFcode, iQuant); if ((iMinSAD < 256) || ((MVequal(*currMV, prevMB->mvs[0])) && - ((uint32_t) iMinSAD < prevMB->sad16))) { + ((int32_t) iMinSAD < prevMB->sad16))) { if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode { if (!MVzero(*currMV)) { @@ -1439,7 +1555,7 @@ if ((iMinSAD <= threshA) || (MVequal(*currMV, prevMB->mvs[0]) && - ((uint32_t) iMinSAD < prevMB->sad16))) { + ((int32_t) iMinSAD < prevMB->sad16))) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast16_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -1466,10 +1582,12 @@ 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, pmv, min_dx, max_dx, + (*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); @@ -1484,7 +1602,7 @@ if (!(MVequal(pmv[0], backupMV))) { iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, pmv, + center_x, center_y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); @@ -1497,8 +1615,9 @@ if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, - max_dy, iEdgedWidth, iDiamondSize, iFcode, + iMinSAD, &newMV, center_x, center_y, + min_dx, max_dx, min_dy, max_dy, + iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); if (iSAD < iMinSAD) { @@ -1516,12 +1635,12 @@ if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step iMinSAD = Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, - iMinSAD, pmv, min_dx, max_dx, min_dy, max_dy, + iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); PMVfast16_Terminate_without_Refine: - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; + currPMV->x = currMV->x - center_x; + currPMV->y = currMV->y - center_y; return iMinSAD; } @@ -1538,11 +1657,12 @@ const uint8_t * const cur, const int x, const int y, - int32_t startx, - int32_t starty, + int32_t start_x, + int32_t start_y, int32_t iMinSAD, VECTOR * const currMV, - const VECTOR * const pmv, + const int center_x, + const int center_y, const int32_t min_dx, const int32_t max_dx, const int32_t min_dy, @@ -1556,11 +1676,12 @@ /* Do a diamond search around given starting point, return SAD of best */ int32_t iDirection = 0; + int32_t iDirectionBackup; int32_t iSAD; VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; + backupMV.x = start_x; + backupMV.y = start_y; /* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ @@ -1569,70 +1690,258 @@ 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 = startx; - currMV->y = starty; + currMV->x = start_x; + currMV->y = start_y; } return iMinSAD; } + + + int32_t -Halfpel8_Refine(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 VECTOR * const pmv, - 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) +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, + int32_t start_x, + 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) { -/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ +/* Do a square search around given starting point, return SAD of best */ + int32_t iDirection = 0; int32_t iSAD; - 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); + backupMV.x = start_x; + backupMV.y = start_y; - return iMinSAD; -} +/* It's one search with full square pattern, and new parts for all following diamonds */ +/* new direction are extra, so 1-4 is normal diamond + 537 + 1*2 + 648 +*/ -#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) + 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 + 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); + + + 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, @@ -1642,6 +1951,8 @@ 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, @@ -1674,7 +1985,7 @@ // const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; - static int32_t threshA, threshB; + int32_t threshA, threshB; int32_t iFound, bPredEq; int32_t iMinSAD, iSAD; @@ -1698,15 +2009,14 @@ } /* 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)) { threshA = 512 / 4; threshB = 1024 / 4; } else { - threshA = psad[0] / 4; /* good estimate */ + threshA = psad[0] / 4; /* good estimate? */ threshB = threshA + 256 / 4; if (threshA < 512 / 4) threshA = 512 / 4; @@ -1728,9 +2038,9 @@ // 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; @@ -1745,11 +2055,11 @@ get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), iEdgedWidth); iMinSAD += - calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, + calc_delta_8(currMV->x - center_x, currMV->y - center_y, (uint8_t) iFcode, iQuant); if ((iMinSAD < 256 / 4) || ((MVequal(*currMV, prevMB->mvs[iSubBlock])) - && ((uint32_t) iMinSAD < + && ((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; @@ -1788,7 +2098,7 @@ // the median prediction might be even better than mv16 if (!MVequal(pmv[0], startMV)) - CHECK_MV8_CANDIDATE(pmv[0].x, pmv[0].y); + CHECK_MV8_CANDIDATE(center_x, center_y); // (0,0) if needed if (!MVzero(pmv[0])) @@ -1804,7 +2114,7 @@ if ((iMinSAD <= threshA) || (MVequal(*currMV, prevMB->mvs[iSubBlock]) && - ((uint32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { + ((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -1863,7 +2173,7 @@ if ((iMinSAD <= threshA) || (MVequal(*currMV, prevMB->mvs[iSubBlock]) && - ((uint32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { + ((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -1885,7 +2195,7 @@ /* 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, pmv, min_dx, max_dx, + currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); @@ -1900,7 +2210,7 @@ if (!(MVequal(pmv[0], backupMV))) { iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, pmv, + 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); @@ -1913,7 +2223,7 @@ if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, + iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); @@ -1932,13 +2242,13 @@ if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step iMinSAD = Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, - iMinSAD, pmv, min_dx, max_dx, min_dy, max_dy, + iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); PMVfast8_Terminate_without_Refine: - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; + currPMV->x = currMV->x - center_x; + currPMV->y = currMV->y - center_y; return iMinSAD; } @@ -1951,6 +2261,10 @@ 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, @@ -1986,7 +2300,7 @@ const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; MACROBLOCK *oldMB = NULL; - static int32_t thresh2; + int32_t thresh2; int32_t bPredEq; int32_t iMinSAD, iSAD = 9999; @@ -2009,7 +2323,7 @@ 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. MinSAD=SAD @@ -2020,7 +2334,9 @@ // Prepare for main loop - *currMV = pmv[0]; /* current best := median prediction */ + currMV->x = start_x; + currMV->y = start_y; + if (!(MotionFlags & PMV_HALFPEL16)) { currMV->x = EVEN(currMV->x); currMV->y = EVEN(currMV->y); @@ -2042,13 +2358,13 @@ get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); iMinSAD += - calc_delta_16(currMV->x - pmv[0].x, currMV->y - pmv[0].y, + calc_delta_16(currMV->x - center_x, currMV->y - center_y, (uint8_t) iFcode, iQuant); // thresh1 is fixed to 256 if ((iMinSAD < 256) || ((MVequal(*currMV, prevMB->mvs[0])) && - ((uint32_t) iMinSAD < prevMB->sad16))) { + ((int32_t) iMinSAD < prevMB->sad16))) { if (MotionFlags & PMV_QUICKSTOP16) goto EPZS16_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -2108,7 +2424,7 @@ if ((iMinSAD <= thresh2) || (MVequal(*currMV, prevMB->mvs[0]) && - ((uint32_t) iMinSAD <= prevMB->sad16))) { + ((int32_t) iMinSAD <= prevMB->sad16))) { if (MotionFlags & PMV_QUICKSTOP16) goto EPZS16_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -2154,10 +2470,10 @@ backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ - if (MotionFlags & PMV_USESQUARES8) + if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = Square16_MainSearch; else - if (MotionFlags & PMV_ADVANCEDDIAMOND8) + if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamond16_MainSearch; else MainSearchPtr = Diamond16_MainSearch; @@ -2166,7 +2482,7 @@ iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, - currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, + currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); if (iSAD < iMinSAD) { @@ -2181,7 +2497,7 @@ if (!(MVequal(pmv[0], backupMV))) { iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, pmv, + pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); } @@ -2194,7 +2510,7 @@ if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, + iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); if (iSAD < iMinSAD) { @@ -2210,15 +2526,15 @@ if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step iMinSAD = Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, - iMinSAD, pmv, min_dx, max_dx, min_dy, max_dy, + iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); EPZS16_Terminate_without_Refine: *oldMB = *prevMB; - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; + currPMV->x = currMV->x - center_x; + currPMV->y = currMV->y - center_y; return iMinSAD; } @@ -2233,6 +2549,8 @@ 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, @@ -2287,7 +2605,7 @@ 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, psad); + bPredEq = get_pmvdata2(pMBs, iWcount, 0, x >> 1, y >> 1, iSubBlock, pmv, psad); /* Step 4: Calculate SAD around the Median prediction. @@ -2322,7 +2640,7 @@ get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), iEdgedWidth); iMinSAD += - calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, + calc_delta_8(currMV->x - center_x, currMV->y - center_y, (uint8_t) iFcode, iQuant); @@ -2398,9 +2716,9 @@ // 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; @@ -2409,7 +2727,7 @@ iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, - currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, + currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); @@ -2425,7 +2743,7 @@ if (!(MVequal(pmv[0], backupMV))) { iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, pmv, + pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); @@ -2438,7 +2756,7 @@ if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { iSAD = (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, + iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); @@ -2455,121 +2773,307 @@ if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step iMinSAD = Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, - iMinSAD, pmv, min_dx, max_dx, min_dy, max_dy, + iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); EPZS8_Terminate_without_Refine: - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; + currPMV->x = currMV->x - center_x; + 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, /* start should be most likely vector */ + const int start_y, + 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, + 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; + + 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 { + + bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); + + threshA = psad[0]; + threshB = threshA + 256; + if (threshA < 512) + threshA = 512; + if (threshA > 1024) + threshA = 1024; + if (threshB > 1792) + threshB = 1792; + + *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); -/* *********************************************************** - bvop motion estimation -// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) -***************************************************************/ - - -void -MotionEstimationBVOP(MBParam * const pParam, - FRAMEINFO * const frame, - // 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 uint32_t mb_width = pParam->mb_width; - const uint32_t mb_height = pParam->mb_height; - const int32_t edged_width = pParam->edged_width; - - uint32_t i, j; - - int32_t f_sad16; - int32_t b_sad16; - int32_t i_sad16; - int32_t d_sad16; - int32_t best_sad; - - VECTOR pmv_dontcare; - - // note: i==horizontal, j==vertical - for (j = 0; j < mb_height; j++) { - 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]; - - 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; - } - - // forward search - f_sad16 = - SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - &frame->image, i, j, frame->motion_flags, - frame->quant, frame->fcode, pParam, f_mbs, - f_mbs /* todo */ , - &mb->mvs[0], &pmv_dontcare); // ignore pmv - - // backward search - b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, &frame->image, i, j, frame->motion_flags, frame->quant, frame->bcode, pParam, b_mbs, b_mbs, /* todo */ - &mb->b_mvs[0], &pmv_dontcare); // ignore pmv - - // interpolate search (simple, but effective) - i_sad16 = - sad16bi_c(frame->image.y + i * 16 + j * 16 * edged_width, - get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - i, j, 16, mb->mvs[0].x, mb->mvs[0].y, - edged_width), get_ref(b_ref->y, b_refH->y, - b_refV->y, b_refHV->y, - i, j, 16, - mb->b_mvs[0].x, - mb->b_mvs[0].x, - edged_width), - edged_width); - - // TODO: direct search - // predictor + range of [-32,32] - 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) { - best_sad = d_sad16; - mb->mode = MODE_DIRECT; + 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; } +