--- trunk/xvidcore/src/motion/motion_est.c 2002/05/07 20:03:18 167 +++ trunk/xvidcore/src/motion/motion_est.c 2002/07/28 17:10:39 348 @@ -1,5 +1,35 @@ /************************************************************************** * + * XVID MPEG-4 VIDEO CODEC + * motion estimation + * + * 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., 675 Mass Ave, Cambridge, MA 02139, USA. + * + *************************************************************************/ + +/************************************************************************** + * * Modifications: * * 01.05.2002 updated MotionEstimationBVOP @@ -47,370 +77,238 @@ #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 - -/* 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 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) - - -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), (int)(1.49591+0.5), (int)(1.68601+0.5), - (int)(1.89187+0.5), (int)(2.11542+0.5), (int)(2.35878+0.5), (int)(2.62429+0.5), (int)(2.91455+0.5), - (int)(3.23253+0.5), (int)(3.58158+0.5), (int)(3.96555+0.5), (int)(4.38887+0.5), (int)(4.85673+0.5), - (int)(5.37519+0.5), (int)(5.95144+0.5), (int)(6.59408+0.5), (int)(7.31349+0.5), (int)(8.12242+0.5), - (int)(9.03669+0.5), (int)(10.0763+0.5), (int)(11.2669+0.5), (int)(12.6426+0.5), (int)(14.2493+0.5), - (int)(16.1512+0.5), (int)(18.442+0.5), (int)(21.2656+0.5), (int)(24.8580+0.5), (int)(29.6436+0.5), - (int)(36.4949+0.5) }; +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), + (int) (1.49591 + 0.5), (int) (1.68601 + 0.5), + (int) (1.89187 + 0.5), (int) (2.11542 + 0.5), (int) (2.35878 + 0.5), + (int) (2.62429 + 0.5), (int) (2.91455 + 0.5), + (int) (3.23253 + 0.5), (int) (3.58158 + 0.5), (int) (3.96555 + 0.5), + (int) (4.38887 + 0.5), (int) (4.85673 + 0.5), + (int) (5.37519 + 0.5), (int) (5.95144 + 0.5), (int) (6.59408 + 0.5), + (int) (7.31349 + 0.5), (int) (8.12242 + 0.5), + (int) (9.03669 + 0.5), (int) (10.0763 + 0.5), (int) (11.2669 + 0.5), + (int) (12.6426 + 0.5), (int) (14.2493 + 0.5), + (int) (16.1512 + 0.5), (int) (18.442 + 0.5), (int) (21.2656 + 0.5), + (int) (24.8580 + 0.5), (int) (29.6436 + 0.5), + (int) (36.4949 + 0.5) +}; -static int32_t *lambda_vec8 = lambda_vec16; /* same table for INTER and INTER4V for now*/ +static int32_t *lambda_vec8 = lambda_vec16; /* same table for INTER and INTER4V for now */ // mv.length table static const uint32_t mvtab[33] = { - 1, 2, 3, 4, 6, 7, 7, 7, - 9, 9, 9, 10, 10, 10, 10, 10, - 10, 10, 10, 10, 10, 10, 10, 10, - 10, 11, 11, 11, 11, 11, 11, 12, 12 + 1, 2, 3, 4, 6, 7, 7, 7, + 9, 9, 9, 10, 10, 10, 10, 10, + 10, 10, 10, 10, 10, 10, 10, 10, + 10, 11, 11, 11, 11, 11, 11, 12, 12 }; -static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) +static __inline uint32_t +mv_bits(int32_t component, + const uint32_t iFcode) { - if (component == 0) + if (component == 0) return 1; - if (component < 0) + if (component < 0) component = -component; - if (iFcode == 1) - { + if (iFcode == 1) { if (component > 32) - component = 32; + component = 32; return mvtab[component] + 1; - } + } - component += (1 << (iFcode - 1)) - 1; - component >>= (iFcode - 1); + component += (1 << (iFcode - 1)) - 1; + component >>= (iFcode - 1); - if (component > 32) + if (component > 32) component = 32; - return mvtab[component] + 1 + iFcode - 1; + return mvtab[component] + 1 + iFcode - 1; } -static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode, const uint32_t iQuant) +static __inline uint32_t +calc_delta_16(const int32_t dx, + const int32_t dy, + const uint32_t iFcode, + const uint32_t iQuant) { - return NEIGH_TEND_16X16 * lambda_vec16[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); + return NEIGH_TEND_16X16 * lambda_vec16[iQuant] * (mv_bits(dx, iFcode) + + mv_bits(dy, iFcode)); } -static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode, const uint32_t iQuant) - +static __inline uint32_t +calc_delta_8(const int32_t dx, + const int32_t dy, + const uint32_t iFcode, + const uint32_t iQuant) { - return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); + 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, - FRAMEINFO * const reference, - const IMAGE * const pRefH, - const IMAGE * const pRefV, - const IMAGE * const pRefHV, - const uint32_t iLimit) - + +bool +MotionEstimation(MBParam * const pParam, + FRAMEINFO * const current, + FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV, + const uint32_t iLimit) { const uint32_t iWcount = pParam->mb_width; const uint32_t iHcount = pParam->mb_height; - MACROBLOCK * pMBs = current->mbs; - IMAGE * pCurrent = ¤t->image; - - MACROBLOCK * prevMBs = reference->mbs; // previous frame - IMAGE * pRef = &reference->image; - - - uint32_t i, j, iIntra = 0; - - VECTOR mv16; - VECTOR pmv16; - - int32_t sad8 = 0; - int32_t sad16; - int32_t deviation; + MACROBLOCK *const pMBs = current->mbs; + MACROBLOCK *const prevMBs = reference->mbs; + const IMAGE *const pCurrent = ¤t->image; + const IMAGE *const pRef = &reference->image; + + static const VECTOR zeroMV = { 0, 0 }; + VECTOR predMV; + + int32_t x, y; + int32_t iIntra = 0; + VECTOR pmv; if (sadInit) - (*sadInit)(); + (*sadInit) (); - // note: i==horizontal, j==vertical - for (i = 0; i < iHcount; i++) - for (j = 0; j < iWcount; j++) - { - MACROBLOCK *pMB = &pMBs[j + i * iWcount]; - MACROBLOCK *prevMB = &prevMBs[j + i * iWcount]; + for (y = 0; y < iHcount; y++) { + for (x = 0; x < iWcount; x ++) { + + MACROBLOCK *const pMB = &pMBs[x + y * iWcount]; - sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - j, i, current->motion_flags, current->quant, current->fcode, - pParam, pMBs, prevMBs, &mv16, &pmv16); - pMB->sad16=sad16; + 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, predMV.x, predMV.y, predMV.x, predMV.y, + current->motion_flags, current->quant, + current->fcode, pParam, pMBs, prevMBs, &pMB->mv16, + &pMB->pmvs[0]); - /* decide: MODE_INTER or MODE_INTRA - if (dev_intra < sad_inter - 2 * nb) use_intra - */ + if (0 < (pMB->sad16 - MV16_INTER_BIAS)) { + int32_t deviation; - deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); - - if (deviation < (sad16 - INTER_BIAS)) - { - pMB->mode = MODE_INTRA; - pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; - pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; - - pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; - - iIntra++; - if(iIntra >= iLimit) - return 1; + deviation = + dev16(pCurrent->y + x * 16 + y * 16 * pParam->edged_width, + pParam->edged_width); - continue; - } + if (deviation < (pMB->sad16 - MV16_INTER_BIAS)) { + pMB->mode = MODE_INTRA; + pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = + pMB->mvs[3] = zeroMV; + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = + pMB->sad8[3] = 0; - if (current->global_flags & XVID_INTER4V) - { - pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - 2 * j, 2 * i, mv16.x, mv16.y, - current->motion_flags, current->quant, current->fcode, - pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); - - pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - 2 * j + 1, 2 * i, mv16.x, mv16.y, - current->motion_flags, current->quant, current->fcode, - pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); - - pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - 2 * j, 2 * i + 1, mv16.x, mv16.y, - current->motion_flags, current->quant, current->fcode, - pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); - - pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - 2 * j + 1, 2 * i + 1, mv16.x, mv16.y, - current->motion_flags, current->quant, current->fcode, - pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); + iIntra++; + if (iIntra >= iLimit) + return 1; - sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; + continue; + } } - - /* decide: MODE_INTER or MODE_INTER4V - mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v - */ + pmv = pMB->pmvs[0]; + if (current->global_flags & XVID_INTER4V) + if ((!(current->global_flags & XVID_LUMIMASKING) || + pMB->dquant == NO_CHANGE)) { + int32_t sad8 = IMV16X16 * current->quant; + + 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, predMV.x, predMV.y, + current->motion_flags, + current->quant, current->fcode, pParam, + pMBs, prevMBs, &pMB->mvs[0], + &pMB->pmvs[0]); + } + 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, predMV.x, predMV.y, + current->motion_flags, + current->quant, current->fcode, pParam, + pMBs, prevMBs, &pMB->mvs[1], + &pMB->pmvs[1]); + } + 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, predMV.x, predMV.y, + current->motion_flags, + current->quant, current->fcode, pParam, + pMBs, prevMBs, &pMB->mvs[2], + &pMB->pmvs[2]); + } + 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, 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 + */ + + if (sad8 < pMB->sad16) { + pMB->mode = MODE_INTER4V; + pMB->sad8[0] *= 4; + pMB->sad8[1] *= 4; + pMB->sad8[2] *= 4; + pMB->sad8[3] *= 4; + continue; + } - if (!(current->global_flags & XVID_LUMIMASKING) || pMB->dquant == NO_CHANGE) - { - if (((current->global_flags & XVID_INTER4V)==0) || - (sad16 < (sad8 + (int32_t)(IMV16X16 * current->quant)))) - { - - sad8 = sad16; - pMB->mode = MODE_INTER; - pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; - pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; - pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad16; - pMB->pmvs[0].x = pmv16.x; - pMB->pmvs[0].y = pmv16.y; - } - else - { - pMB->mode = MODE_INTER4V; - pMB->sad8[0] *= 4; - pMB->sad8[1] *= 4; - pMB->sad8[2] *= 4; - pMB->sad8[3] *= 4; } - } - else - { - sad8 = sad16; - pMB->mode = MODE_INTER; - pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; - pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; - pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad16; - pMB->pmvs[0].x = pmv16.x; - pMB->pmvs[0].y = pmv16.y; + pMB->mode = MODE_INTER; + pMB->pmvs[0] = pmv; /* pMB->pmvs[1] = pMB->pmvs[2] = pMB->pmvs[3] are not needed for INTER */ + 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 MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) - -#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) - #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); } \ } @@ -420,7 +318,7 @@ && ((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); } } \ } @@ -430,7 +328,7 @@ && ((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); } } \ } @@ -440,7 +338,7 @@ && ((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; } } \ } @@ -448,15 +346,15 @@ #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; } \ } - + #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); } \ } @@ -466,7 +364,7 @@ && ((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); } } \ } @@ -476,7 +374,7 @@ && ((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); } } \ } @@ -486,7 +384,7 @@ && ((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; } } \ } @@ -512,9 +410,10 @@ 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), @@ -524,98 +423,114 @@ currMV->x = 0; currMV->y = 0; - currPMV->x = -pred_x; - currPMV->y = -pred_y; + currPMV->x = -pred.x; + currPMV->y = -pred.y; return iSAD; } */ -int32_t Diamond16_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 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) +int32_t +Diamond16_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 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, + 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 diamond search around given starting point, return SAD of best */ - int32_t iDirection=0; + 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 */ - 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 = startx; - currMV->y = starty; + 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; + iDirectionBackup = iDirection; + + if (iDirectionBackup != 2) + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y, 1); + if (iDirectionBackup != 1) + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y, 2); + if (iDirectionBackup != 4) + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, + backupMV.y - iDiamondSize, 3); + if (iDirectionBackup != 3) + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, + backupMV.y + iDiamondSize, 4); + } + } else { + currMV->x = start_x; + currMV->y = start_y; } return iMinSAD; } -int32_t Square16_MainSearch( - const uint8_t * const pRef, +int32_t +Square16_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 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 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 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 square search around given starting point, return SAD of best */ - int32_t iDirection=0; + int32_t iDirection = 0; 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 */ /* new direction are extra, so 1-4 is normal diamond @@ -624,351 +539,986 @@ 648 */ - 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); - - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - + 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); + + CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + + + if (iDirection) { + while (!iFound) { + iFound = 1; + backupMV = *currMV; + + switch (iDirection) { + case 1: + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y, 1); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + break; + case 2: + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, + 2); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + + case 3: + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, + 4); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + + case 4: + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, + 3); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + break; + + case 5: + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, + 1); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, + 3); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + break; + + case 6: + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, + 2); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, + 3); + + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + + break; + + case 7: + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y, 1); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, + 4); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + + case 8: + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, + 2); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, + 4); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + default: + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, + 1); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, + 2); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, + 3); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, + 4); + + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y - iDiamondSize, 5); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y + iDiamondSize, 6); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y - iDiamondSize, 7); + CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y + iDiamondSize, 8); + break; + } + } + } else { + currMV->x = start_x; + currMV->y = start_y; + } + return iMinSAD; +} - if (iDirection) - while (!iFound) - { - iFound = 1; - backupMV=*currMV; - - switch (iDirection) + +int32_t +Full16_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 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, + const int32_t max_dy, + const int32_t iEdgedWidth, + const int32_t iDiamondSize, + const int32_t iFcode, + const int32_t iQuant, + int iFound) +{ + int32_t iSAD; + int32_t dx, dy; + VECTOR backupMV; + + 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) + NOCHECK_MV16_CANDIDATE(dx, dy); + + return iMinSAD; +} + +int32_t +AdvDiamond16_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, + 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, + const int32_t max_dy, + const int32_t iEdgedWidth, + const int32_t iDiamondSize, + const int32_t iFcode, + const int32_t iQuant, + int iDirection) +{ + + int32_t iSAD; + +/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ + + if (iDirection) { + 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(start_x - iDiamondSize, start_y, 1); + + if (bDirection & 2) + CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); + + if (bDirection & 4) + CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); + + if (bDirection & 8) + CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); + + /* now we're doing diagonal checks near our candidate */ + + if (iDirection) //checking if anything found { - case 1: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - break; + bDirection = iDirection; + iDirection = 0; + start_x = currMV->x; + start_y = currMV->y; + if (bDirection & 3) //our candidate is left or right + { + 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(start_x + iDiamondSize, start_y, 2); + CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); + } + + if (iDirection) { + bDirection += iDirection; + start_x = currMV->x; + start_y = currMV->y; + } + } else //about to quit, eh? not so fast.... + { + switch (bDirection) { case 2: - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,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 3: - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); + case 1: + + 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(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(backupMV.x,backupMV.y-iDiamondSize,3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); + 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 5: - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - break; - - case 6: - 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-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - + case 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 7: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); + case 1 + 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 8: - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - break; - default: - 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); - - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); + case 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(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(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; + start_x = currMV->x; + start_y = currMV->y; + } } } - else - { - currMV->x = startx; - currMV->y = starty; - } + while (1); //forever + } 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, -int32_t Full16_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 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) { +/* Do a diamond search around given starting point, return SAD of best */ + int32_t iSAD; - int32_t dx,dy; - VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; + + 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; - for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) - for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) - NOCHECK_MV16_CANDIDATE(dx,dy); + 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; } -int32_t Full8_MainSearch( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, +/* 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 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 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; - int32_t dx,dy; + VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; + + 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; - for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) - for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) - NOCHECK_MV8_CANDIDATE(dx,dy); + 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 Halfpel16_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) +int32_t +AdvDiamond8_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, + 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, + const int32_t max_dy, + const int32_t iEdgedWidth, + const int32_t iDiamondSize, + const int32_t iFcode, + const int32_t iQuant, + int iDirection) { -/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ int32_t iSAD; - VECTOR backupMV = *currMV; - - CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); - CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); - CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); - CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); - CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y); - CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y+1); - CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y+1); - CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y+1); - - return iMinSAD; -} -#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) +/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ + if (iDirection) { + 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; -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) + 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(start_x - iDiamondSize, start_y, 1); + + if (bDirection & 2) + CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); + + if (bDirection & 4) + CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); + + if (bDirection & 8) + CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); + + /* now we're doing diagonal checks near our candidate */ + + if (iDirection) //checking if anything found + { + bDirection = iDirection; + iDirection = 0; + start_x = currMV->x; + start_y = currMV->y; + if (bDirection & 3) //our candidate is left or right + { + 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(start_x + iDiamondSize, start_y, 2); + CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); + } + + if (iDirection) { + bDirection += iDirection; + start_x = currMV->x; + start_y = currMV->y; + } + } else //about to quit, eh? not so fast.... + { + switch (bDirection) { + case 2: + 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(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(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(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(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(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(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(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(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; + start_x = currMV->x; + start_y = currMV->y; + } + } + } + while (1); //forever + } + return iMinSAD; +} + + +int32_t +Full8_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 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, + const int32_t max_dy, + const int32_t iEdgedWidth, + const int32_t iDiamondSize, + const int32_t iFcode, + const int32_t iQuant, + int iFound) +{ + int32_t iSAD; + int32_t dx, dy; + VECTOR backupMV; + + 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) + NOCHECK_MV8_CANDIDATE(dx, dy); + + return iMinSAD; +} + +Halfpel8_RefineFuncPtr Halfpel8_Refine; + +int32_t +Halfpel16_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 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_MV16_CANDIDATE(backupMV.x - 1, backupMV.y - 1); + CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y - 1); + CHECK_MV16_CANDIDATE(backupMV.x + 1, backupMV.y - 1); + CHECK_MV16_CANDIDATE(backupMV.x - 1, backupMV.y); + CHECK_MV16_CANDIDATE(backupMV.x + 1, backupMV.y); + CHECK_MV16_CANDIDATE(backupMV.x - 1, backupMV.y + 1); + CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y + 1); + CHECK_MV16_CANDIDATE(backupMV.x + 1, backupMV.y + 1); + + return iMinSAD; +} + +#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) + + + +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 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, + 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 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 int32_t iEdgedWidth = pParam->edged_width; - const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; + const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; 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 backupMV; /* just for PMVFAST */ + VECTOR pmv[4]; int32_t psad[4]; - - const MACROBLOCK * const pMB = pMBs + x + y * iWcount; - const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; - static int32_t threshA,threshB; - int32_t bPredEq; - int32_t iMinSAD,iSAD; + MainSearch16FuncPtr MainSearchPtr; + + const MACROBLOCK *const prevMB = prevMBs + 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); + 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 (!(MotionFlags & PMV_HALFPEL16 )) - { min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - + if (!(MotionFlags & PMV_HALFPEL16)) { + min_dx = EVEN(min_dx); + max_dx = EVEN(max_dx); + min_dy = EVEN(min_dy); + max_dy = EVEN(max_dy); + } - bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); + /* 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; + if ((x == 0) && (y == 0)) { + threshA = 512; threshB = 1024; - - } - else - { + } else { threshA = psad[0]; - threshB = threshA+256; - if (threshA< 512) threshA = 512; - if (threshA>1024) threshA = 1024; - if (threshB>1792) threshB = 1792; + threshB = threshA + 256; + if (threshA < 512) + threshA = 512; + if (threshA > 1024) + threshA = 1024; + if (threshB > 1792) + threshB = 1792; } - iFound=0; - -/* Step 2: 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->mvs[0]) ) ) - iFound=2; - -/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ - - if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) - iDiamondSize=1; // halfpel! - else - iDiamondSize=2; // halfpel! - - if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) - iDiamondSize*=2; + iFound = 0; /* Step 4: Calculate SAD around the Median prediction. MinSAD=SAD If Motion Vector equal to Previous frame motion vector and MinSADx = start_x; + currMV->y = start_y; - *currMV=pmv[0]; /* current best := prediction */ - if (!(MotionFlags & PMV_HALFPEL16 )) - { /* This should NOT be necessary! */ + if (!(MotionFlags & PMV_HALFPEL16)) { /* This should NOT be necessary! */ currMV->x = EVEN(currMV->x); currMV->y = EVEN(currMV->y); } - - if (currMV->x > max_dx) - { - currMV->x=max_dx; - } - if (currMV->x < min_dx) - { - currMV->x=min_dx; + + if (currMV->x > max_dx) { + currMV->x = max_dx; } - if (currMV->y > max_dy) - { - currMV->y=max_dy; + if (currMV->x < min_dx) { + currMV->x = min_dx; } - if (currMV->y < min_dy) - { - currMV->y=min_dy; + if (currMV->y > max_dy) { + currMV->y = max_dy; } - - iMinSAD = sad16( cur, - 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, (uint8_t)iFcode, iQuant); - - if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) - { - - if (MotionFlags & PMV_QUICKSTOP16) + if (currMV->y < min_dy) { + currMV->y = min_dy; + } + + iMinSAD = + sad16(cur, + get_ref_mv(pRef, pRefH, pRefV, pRefHV, 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->mvs[0])) && + ((int32_t) iMinSAD < prevMB->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_QUICKSTOP16) goto PMVfast16_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) goto PMVfast16_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->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 = 1; // halfpel! + else + iDiamondSize = 2; // halfpel! + + if (!(MotionFlags & PMV_HALFPELDIAMOND16)) + iDiamondSize *= 2; + /* 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. @@ -978,52 +1528,66 @@ // (0,0) is always possible - CHECK_MV16_ZERO; + if (!MVzero(pmv[0])) + CHECK_MV16_ZERO; // previous frame MV is always possible - CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); - + + if (!MVzero(prevMB->mvs[0])) + if (!MVequal(prevMB->mvs[0], pmv[0])) + CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); + // left neighbour, if allowed - if (x != 0) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[1].x = EVEN(pmv[1].x); - pmv[1].y = EVEN(pmv[1].y); - } - CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); - } + if (!MVzero(pmv[1])) + if (!MVequal(pmv[1], prevMB->mvs[0])) + if (!MVequal(pmv[1], pmv[0])) { + if (!(MotionFlags & PMV_HALFPEL16)) { + pmv[1].x = EVEN(pmv[1].x); + pmv[1].y = EVEN(pmv[1].y); + } + + CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); + } // top neighbour, if allowed - if (y != 0) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[2].x = EVEN(pmv[2].x); - pmv[2].y = EVEN(pmv[2].y); - } - CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); - + if (!MVzero(pmv[2])) + if (!MVequal(pmv[2], prevMB->mvs[0])) + if (!MVequal(pmv[2], pmv[0])) + if (!MVequal(pmv[2], pmv[1])) { + if (!(MotionFlags & PMV_HALFPEL16)) { + pmv[2].x = EVEN(pmv[2].x); + pmv[2].y = EVEN(pmv[2].y); + } + CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); + // top right neighbour, if allowed - if ((uint32_t)x != (iWcount-1)) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[3].x = EVEN(pmv[3].x); - pmv[3].y = EVEN(pmv[3].y); - } - CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); - } - } - - if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96)*/ ) + if (!MVzero(pmv[3])) + if (!MVequal(pmv[3], prevMB->mvs[0])) + if (!MVequal(pmv[3], pmv[0])) + if (!MVequal(pmv[3], pmv[1])) + if (!MVequal(pmv[3], pmv[2])) { + if (!(MotionFlags & PMV_HALFPEL16)) { + pmv[3].x = EVEN(pmv[3].x); + pmv[3].y = EVEN(pmv[3].y); + } + 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 MinSADmvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) + if ((iMinSAD <= threshA) || + (MVequal(*currMV, prevMB->mvs[0]) && + ((int32_t) iMinSAD < prevMB->sad16))) { + if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast16_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) goto PMVfast16_Terminate_with_Refine; @@ -1040,48 +1604,57 @@ Refine by using small diamond and goto step 10. */ - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + 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 = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV->x, currMV->y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { + 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) - { + if (MotionFlags & PMV_EXTSEARCH16) { /* extended: search (up to) two more times: orignal prediction and (0,0) */ - if (!(MVequal(pmv[0],backupMV)) ) - { iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } - } + if (!(MVequal(pmv[0], backupMV))) { + iSAD = + (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, + center_x, center_y, iMinSAD, &newMV, center_x, center_y, + min_dx, max_dx, min_dy, max_dy, iEdgedWidth, + iDiamondSize, iFcode, iQuant, iFound); - if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) - { iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - 0, 0, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; + 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; + } } } @@ -1089,16 +1662,16 @@ Step 10: The motion vector is chosen according to the block corresponding to MinSAD. */ -PMVfast16_Terminate_with_Refine: - 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, iFcode, iQuant, iEdgedWidth); - -PMVfast16_Terminate_without_Refine: - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; + PMVfast16_Terminate_with_Refine: + 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); + + PMVfast16_Terminate_without_Refine: + currPMV->x = currMV->x - center_x; + currPMV->y = currMV->y - center_y; return iMinSAD; } @@ -1107,122 +1680,325 @@ -int32_t Diamond8_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 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) +int32_t +Diamond8_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 diamond search around given starting point, return SAD of best */ - int32_t iDirection=0; + 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 */ - 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); - - if (iDirection) - while (!iFound) - { - iFound = 1; - backupMV=*currMV; // since iDirection!=0, this is well defined! - - if ( iDirection != 2) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); - if ( iDirection != 1) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); - if ( iDirection != 4) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); - if ( iDirection != 3) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); + 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); + + if (iDirection) { + while (!iFound) { + iFound = 1; + backupMV = *currMV; // since iDirection!=0, this is well defined! + iDirectionBackup = iDirection; + + if (iDirectionBackup != 2) + CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, + backupMV.y, 1); + if (iDirectionBackup != 1) + CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, + backupMV.y, 2); + if (iDirectionBackup != 4) + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, + backupMV.y - iDiamondSize, 3); + if (iDirectionBackup != 3) + CHECK_MV8_CANDIDATE_FOUND(backupMV.x, + backupMV.y + iDiamondSize, 4); + } + } else { + currMV->x = start_x; + currMV->y = start_y; + } + return iMinSAD; +} + + + + +int32_t +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 square search around given starting point, return SAD of best */ + + int32_t iDirection = 0; + int32_t iSAD; + VECTOR backupMV; + + backupMV.x = start_x; + backupMV.y = start_y; + +/* 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 +*/ + + 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); + + 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 = startx; - currMV->y = starty; + } else { + 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) + + + + +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); - + + 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 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 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 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 int32_t iEdgedWidth = pParam->edged_width; - const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; + const uint8_t *cur = pCur->y + x * 8 + y * 8 * iEdgedWidth; int32_t iDiamondSize; @@ -1230,98 +2006,119 @@ int32_t max_dx; int32_t min_dy; int32_t max_dy; - + VECTOR pmv[4]; int32_t psad[4]; VECTOR newMV; VECTOR backupMV; - - const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; - const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; + VECTOR startMV; - static int32_t threshA,threshB; - int32_t iFound,bPredEq; - int32_t iMinSAD,iSAD; +// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; + const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; - int32_t iSubBlock = ((y&1)<<1) + (x&1); + int32_t threshA, threshB; + int32_t iFound, bPredEq; + int32_t iMinSAD, iSAD; -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, - x, y, 8, iWidth, iHeight, iFcode); + int32_t iSubBlock = (y & 1) + (y & 1) + (x & 1); - if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) - { min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ - + MainSearch8FuncPtr MainSearchPtr; - bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); + /* Init variables */ + startMV.x = start_x; + startMV.y = start_y; - if ((x==0) && (y==0) ) - { - threshA = 512/4; - threshB = 1024/4; - - } - else - { - threshA = psad[0]/4; /* good estimate */ - threshB = threshA+256/4; - if (threshA< 512/4) threshA = 512/4; - if (threshA>1024/4) threshA = 1024/4; - if (threshB>1792/4) threshB = 1792/4; + /* Get maximum range */ + get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 8, iWidth, iHeight, + iFcode); + + if (!(MotionFlags & PMV_HALFPELDIAMOND8)) { + min_dx = EVEN(min_dx); + max_dx = EVEN(max_dx); + min_dy = EVEN(min_dy); + max_dy = EVEN(max_dy); } - iFound=0; - -/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion - vector of the median. - If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 -*/ + /* 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 ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[iSubBlock]) ) ) - iFound=2; + if ((x == 0) && (y == 0)) { + threshA = 512 / 4; + threshB = 1024 / 4; -/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ - - if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) - iDiamondSize=1; // 1 halfpel! - else - iDiamondSize=2; // 2 halfpel = 1 full pixel! + } else { + threshA = psad[0] / 4; /* good estimate? */ + threshB = threshA + 256 / 4; + if (threshA < 512 / 4) + threshA = 512 / 4; + if (threshA > 1024 / 4) + threshA = 1024 / 4; + if (threshB > 1792 / 4) + threshB = 1792 / 4; + } - if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) - iDiamondSize*=2; + iFound = 0; /* Step 4: Calculate SAD around the Median prediction. MinSAD=SAD If Motion Vector equal to Previous frame motion vector and MinSADx=start_x; /* start with mv16 */ - currMV->y=start_y; - - iMinSAD = sad8( cur, - 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, (uint8_t)iFcode, iQuant); - - if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,prevMB->mvs[iSubBlock])) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) + if (MotionFlags & PMV_USESQUARES8) + MainSearchPtr = Square8_MainSearch; + else + + if (MotionFlags & PMV_ADVANCEDDIAMOND8) + MainSearchPtr = AdvDiamond8_MainSearch; + else + MainSearchPtr = Diamond8_MainSearch; + + + *currMV = startMV; + + iMinSAD = + sad8(cur, + get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, + iEdgedWidth), iEdgedWidth); + iMinSAD += + calc_delta_8(currMV->x - center_x, currMV->y - center_y, + (uint8_t) iFcode, iQuant); + + if ((iMinSAD < 256 / 4) || ((MVequal(*currMV, prevMB->mvs[iSubBlock])) + && ((int32_t) iMinSAD < + prevMB->sad8[iSubBlock]))) { + if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) goto PMVfast8_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->mvs[iSubBlock]))) + 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 / 4) || (bPredEq)) + iDiamondSize = 1; // 1 halfpel! + else + iDiamondSize = 2; // 2 halfpel = 1 full pixel! + + if (!(MotionFlags & PMV_HALFPELDIAMOND8)) + iDiamondSize *= 2; + /* Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. @@ -1330,46 +2127,75 @@ If MV is (0,0) subtract offset. */ -// the prediction might be even better than mv16 - CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); +// the median prediction might be even better than mv16 -// (0,0) is always possible - CHECK_MV8_ZERO; + if (!MVequal(pmv[0], startMV)) + CHECK_MV8_CANDIDATE(center_x, center_y); -// previous frame MV is always possible - CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); - -// left neighbour, if allowed - if (psad[1] != MV_MAX_ERROR) - { - if (!(MotionFlags & PMV_HALFPEL8 )) - { pmv[1].x = EVEN(pmv[1].x); - pmv[1].y = EVEN(pmv[1].y); - } - CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); +// (0,0) if needed + if (!MVzero(pmv[0])) + if (!MVzero(startMV)) + CHECK_MV8_ZERO; + +// previous frame MV if needed + if (!MVzero(prevMB->mvs[iSubBlock])) + if (!MVequal(prevMB->mvs[iSubBlock], startMV)) + if (!MVequal(prevMB->mvs[iSubBlock], pmv[0])) + CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, + prevMB->mvs[iSubBlock].y); + + if ((iMinSAD <= threshA) || + (MVequal(*currMV, prevMB->mvs[iSubBlock]) && + ((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { + if (MotionFlags & PMV_QUICKSTOP16) + goto PMVfast8_Terminate_without_Refine; + if (MotionFlags & PMV_EARLYSTOP16) + goto PMVfast8_Terminate_with_Refine; } -// top neighbour, if allowed - if (psad[2] != MV_MAX_ERROR) - { - if (!(MotionFlags & PMV_HALFPEL8 )) - { pmv[2].x = EVEN(pmv[2].x); - pmv[2].y = EVEN(pmv[2].y); - } - CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); - -// top right neighbour, if allowed - if (psad[3] != MV_MAX_ERROR) - { - if (!(MotionFlags & PMV_HALFPEL8 )) - { pmv[3].x = EVEN(pmv[3].x); - pmv[3].y = EVEN(pmv[3].y); - } - CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); - } - } +// left neighbour, if allowed and needed + if (!MVzero(pmv[1])) + if (!MVequal(pmv[1], startMV)) + if (!MVequal(pmv[1], prevMB->mvs[iSubBlock])) + if (!MVequal(pmv[1], pmv[0])) { + if (!(MotionFlags & PMV_HALFPEL8)) { + pmv[1].x = EVEN(pmv[1].x); + pmv[1].y = EVEN(pmv[1].y); + } + CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); + } +// top neighbour, if allowed and needed + if (!MVzero(pmv[2])) + if (!MVequal(pmv[2], startMV)) + if (!MVequal(pmv[2], prevMB->mvs[iSubBlock])) + if (!MVequal(pmv[2], pmv[0])) + if (!MVequal(pmv[2], pmv[1])) { + if (!(MotionFlags & PMV_HALFPEL8)) { + pmv[2].x = EVEN(pmv[2].x); + pmv[2].y = EVEN(pmv[2].y); + } + CHECK_MV8_CANDIDATE(pmv[2].x, pmv[2].y); + +// top right neighbour, if allowed and needed + if (!MVzero(pmv[3])) + if (!MVequal(pmv[3], startMV)) + if (!MVequal(pmv[3], prevMB->mvs[iSubBlock])) + if (!MVequal(pmv[3], pmv[0])) + if (!MVequal(pmv[3], pmv[1])) + if (!MVequal(pmv[3], pmv[2])) { + if (! + (MotionFlags & + PMV_HALFPEL8)) { + pmv[3].x = EVEN(pmv[3].x); + pmv[3].y = EVEN(pmv[3].y); + } + CHECK_MV8_CANDIDATE(pmv[3].x, + pmv[3].y); + } + } - if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) + if ((MVzero(*currMV)) && + (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) iMinSAD -= MV8_00_BIAS; @@ -1377,9 +2203,10 @@ If Motion Vector equal to Previous frame motion vector and MinSADmvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) + if ((iMinSAD <= threshA) || + (MVequal(*currMV, prevMB->mvs[iSubBlock]) && + ((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { + if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) goto PMVfast8_Terminate_with_Refine; @@ -1395,228 +2222,232 @@ Refine by using small diamond and goto step 10. */ - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ /* default: use best prediction as starting point for one call of PMVfast_MainSearch */ - iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV->x, currMV->y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { + 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_EXTSEARCH8) - { + if (MotionFlags & PMV_EXTSEARCH8) { /* extended: search (up to) two more times: orignal prediction and (0,0) */ - if (!(MVequal(pmv[0],backupMV)) ) - { iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } - } + 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 ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) - { iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - 0, 0, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; + 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. By performing an optional local half-pixel search, we can refine this result even further. */ - -PMVfast8_Terminate_with_Refine: - 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, iFcode, iQuant, iEdgedWidth); + PMVfast8_Terminate_with_Refine: + if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step + iMinSAD = + Halfpel8_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); + + + PMVfast8_Terminate_without_Refine: + currPMV->x = currMV->x - center_x; + currPMV->y = currMV->y - center_y; -PMVfast8_Terminate_without_Refine: - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; - return iMinSAD; } -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 +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 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 uint32_t iHcount = pParam->mb_height; + const uint32_t iWcount = pParam->mb_width; + const uint32_t iHcount = pParam->mb_height; const int32_t iWidth = pParam->width; const int32_t iHeight = pParam->height; - const int32_t iEdgedWidth = pParam->edged_width; + const int32_t iEdgedWidth = pParam->edged_width; - const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; + const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; int32_t min_dx; int32_t max_dx; int32_t min_dy; int32_t max_dy; - + VECTOR newMV; VECTOR backupMV; - + VECTOR pmv[4]; int32_t psad[8]; - - static MACROBLOCK * oldMBs = NULL; - const MACROBLOCK * const pMB = pMBs + x + y * iWcount; - const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; - MACROBLOCK * oldMB = NULL; - - static int32_t thresh2; - int32_t bPredEq; - int32_t iMinSAD,iSAD=9999; - - MainSearch16FuncPtr EPZSMainSearchPtr; - - if (oldMBs == NULL) - { oldMBs = (MACROBLOCK*) calloc(iWcount*iHcount,sizeof(MACROBLOCK)); -// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); + + static MACROBLOCK *oldMBs = NULL; + +// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; + const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; + MACROBLOCK *oldMB = NULL; + + int32_t thresh2; + int32_t bPredEq; + int32_t iMinSAD, iSAD = 9999; + + MainSearch16FuncPtr MainSearchPtr; + + if (oldMBs == NULL) { + oldMBs = (MACROBLOCK *) calloc(iWcount * iHcount, sizeof(MACROBLOCK)); +// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); } oldMB = oldMBs + x + y * iWcount; /* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, - x, y, 16, iWidth, iHeight, iFcode); + get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, + iFcode); - if (!(MotionFlags & PMV_HALFPEL16 )) - { min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - 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); + if (!(MotionFlags & PMV_HALFPEL16)) { + min_dx = EVEN(min_dx); + max_dx = EVEN(max_dx); + min_dy = EVEN(min_dy); + 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 If Motion Vector equal to Previous frame motion vector and MinSADx = start_x; + currMV->y = start_y; + + if (!(MotionFlags & PMV_HALFPEL16)) { currMV->x = EVEN(currMV->x); currMV->y = EVEN(currMV->y); } - - if (currMV->x > max_dx) - currMV->x=max_dx; - if (currMV->x < min_dx) - currMV->x=min_dx; - if (currMV->y > max_dy) - currMV->y=max_dy; - if (currMV->y < min_dy) - currMV->y=min_dy; -/***************** This is predictor SET A: only median prediction ******************/ - - iMinSAD = sad16( cur, - 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, (uint8_t)iFcode, iQuant); - + if (currMV->x > max_dx) + currMV->x = max_dx; + if (currMV->x < min_dx) + currMV->x = min_dx; + if (currMV->y > max_dy) + currMV->y = max_dy; + if (currMV->y < min_dy) + currMV->y = min_dy; + +/***************** This is predictor SET A: only median prediction ******************/ + + iMinSAD = + sad16(cur, + get_ref_mv(pRef, pRefH, pRefV, pRefHV, 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); + // thresh1 is fixed to 256 - if ( (iMinSAD < 256 ) || ( (MVequal(*currMV, prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) - goto EPZS16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto EPZS16_Terminate_with_Refine; - } + if ((iMinSAD < 256) || + ((MVequal(*currMV, prevMB->mvs[0])) && + ((int32_t) iMinSAD < prevMB->sad16))) { + if (MotionFlags & PMV_QUICKSTOP16) + goto EPZS16_Terminate_without_Refine; + if (MotionFlags & PMV_EARLYSTOP16) + goto EPZS16_Terminate_with_Refine; + } -/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ +/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ // previous frame MV - CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); + CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); // set threshhold based on Min of Prediction and SAD of collocated block // CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want - if ((x==0) && (y==0) ) - { - thresh2 = 512; - } - else - { + if ((x == 0) && (y == 0)) { + thresh2 = 512; + } else { /* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ - thresh2 = MIN(psad[0],iSAD)*6/5 + 128; + thresh2 = MIN(psad[0], iSAD) * 6 / 5 + 128; } // MV=(0,0) is often a good choice CHECK_MV16_ZERO; - + // left neighbour, if allowed - if (x != 0) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[1].x = EVEN(pmv[1].x); + if (x != 0) { + if (!(MotionFlags & PMV_HALFPEL16)) { + pmv[1].x = EVEN(pmv[1].x); pmv[1].y = EVEN(pmv[1].y); } - CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); + CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); } - // top neighbour, if allowed - if (y != 0) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[2].x = EVEN(pmv[2].x); + if (y != 0) { + if (!(MotionFlags & PMV_HALFPEL16)) { + pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); } - CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); - + CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); + // top right neighbour, if allowed - if ((uint32_t)x != (iWcount-1)) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[3].x = EVEN(pmv[3].x); + if ((uint32_t) x != (iWcount - 1)) { + if (!(MotionFlags & PMV_HALFPEL16)) { + pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); } - CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); + CHECK_MV16_CANDIDATE(pmv[3].x, pmv[3].y); } } @@ -1624,99 +2455,98 @@ Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] */ - if ( (iMinSAD <= thresh2) - || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) - goto EPZS16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto EPZS16_Terminate_with_Refine; - } + if ((iMinSAD <= thresh2) + || (MVequal(*currMV, prevMB->mvs[0]) && + ((int32_t) iMinSAD <= prevMB->sad16))) { + if (MotionFlags & PMV_QUICKSTOP16) + goto EPZS16_Terminate_without_Refine; + if (MotionFlags & PMV_EARLYSTOP16) + goto EPZS16_Terminate_with_Refine; + } /***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ - backupMV = prevMB->mvs[0]; // collocated MV - backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X - backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y + backupMV = prevMB->mvs[0]; // collocated MV + backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x); // acceleration X + backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y); // acceleration Y - CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); + CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y); // left neighbour - if (x != 0) - CHECK_MV16_CANDIDATE((prevMB-1)->mvs[0].x,(prevMB-1)->mvs[0].y); + if (x != 0) + CHECK_MV16_CANDIDATE((prevMB - 1)->mvs[0].x, (prevMB - 1)->mvs[0].y); // top neighbour if (y != 0) - CHECK_MV16_CANDIDATE((prevMB-iWcount)->mvs[0].x,(prevMB-iWcount)->mvs[0].y); + CHECK_MV16_CANDIDATE((prevMB - iWcount)->mvs[0].x, + (prevMB - iWcount)->mvs[0].y); // right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs - if ((uint32_t)x != iWcount-1) - CHECK_MV16_CANDIDATE((prevMB+1)->mvs[0].x,(prevMB+1)->mvs[0].y); + if ((uint32_t) x != iWcount - 1) + CHECK_MV16_CANDIDATE((prevMB + 1)->mvs[0].x, (prevMB + 1)->mvs[0].y); // bottom neighbour, dito - if ((uint32_t)y != iHcount-1) - CHECK_MV16_CANDIDATE((prevMB+iWcount)->mvs[0].x,(prevMB+iWcount)->mvs[0].y); + if ((uint32_t) y != iHcount - 1) + CHECK_MV16_CANDIDATE((prevMB + iWcount)->mvs[0].x, + (prevMB + iWcount)->mvs[0].y); /* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ - if (iMinSAD <= thresh2) - { - if (MotionFlags & PMV_QUICKSTOP16) - goto EPZS16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto EPZS16_Terminate_with_Refine; - } + if (iMinSAD <= thresh2) { + if (MotionFlags & PMV_QUICKSTOP16) + goto EPZS16_Terminate_without_Refine; + if (MotionFlags & PMV_EARLYSTOP16) + goto EPZS16_Terminate_with_Refine; + } /************ (if Diamond Search) **************/ - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ - -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ + backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ if (MotionFlags & PMV_USESQUARES16) - EPZSMainSearchPtr = Square16_MainSearch; + MainSearchPtr = Square16_MainSearch; + else + if (MotionFlags & PMV_ADVANCEDDIAMOND16) + MainSearchPtr = AdvDiamond16_MainSearch; else - EPZSMainSearchPtr = Diamond16_MainSearch; + MainSearchPtr = Diamond16_MainSearch; - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - 2, iFcode, iQuant, 0); - - if (iSAD < iMinSAD) - { +/* 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, 2, iFcode, iQuant, 0); + + if (iSAD < iMinSAD) { *currMV = newMV; iMinSAD = iSAD; } - if (MotionFlags & PMV_EXTSEARCH16) - { + if (MotionFlags & PMV_EXTSEARCH16) { /* extended mode: search (up to) two more times: orignal prediction and (0,0) */ - if (!(MVequal(pmv[0],backupMV)) ) - { - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 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, + 2, iFcode, iQuant, 0); } - - if (iSAD < iMinSAD) - { + + if (iSAD < iMinSAD) { *currMV = newMV; iMinSAD = iSAD; } - - if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) - { - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - 0, 0, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); - - if (iSAD < iMinSAD) - { + + 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, 2, iFcode, iQuant, 0); + + if (iSAD < iMinSAD) { *currMV = newMV; iMinSAD = iSAD; } @@ -1725,86 +2555,91 @@ /*************** Choose best MV found **************/ -EPZS16_Terminate_with_Refine: - 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, iFcode, iQuant, iEdgedWidth); + EPZS16_Terminate_with_Refine: + 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); -EPZS16_Terminate_without_Refine: + 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; } -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) +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 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) { /* Please not that EPZS might not be a good choice for 8x8-block motion search ! */ 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 int32_t iEdgedWidth = pParam->edged_width; - const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; + const uint8_t *cur = pCur->y + x * 8 + y * 8 * iEdgedWidth; + + int32_t iDiamondSize = 1; - int32_t iDiamondSize=1; - int32_t min_dx; int32_t max_dx; int32_t min_dy; int32_t max_dy; - + VECTOR newMV; VECTOR backupMV; - + VECTOR pmv[4]; int32_t psad[8]; - const int32_t iSubBlock = ((y&1)<<1) + (x&1); - - const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; - const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; + const int32_t iSubBlock = ((y & 1) << 1) + (x & 1); + +// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; + const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; - int32_t bPredEq; - int32_t iMinSAD,iSAD=9999; + int32_t bPredEq; + int32_t iMinSAD, iSAD = 9999; - MainSearch8FuncPtr EPZSMainSearchPtr; + MainSearch8FuncPtr MainSearchPtr; /* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, - x, y, 8, iWidth, iHeight, iFcode); + get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 8, iWidth, iHeight, + iFcode); /* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ - if (!(MotionFlags & PMV_HALFPEL8 )) - { min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - 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); + if (!(MotionFlags & PMV_HALFPEL8)) { + min_dx = EVEN(min_dx); + max_dx = EVEN(max_dx); + min_dy = EVEN(min_dy); + 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); /* Step 4: Calculate SAD around the Median prediction. @@ -1816,76 +2651,73 @@ // Prepare for main loop - - if (!(MotionFlags & PMV_HALFPEL8)) - { + + if (!(MotionFlags & PMV_HALFPEL8)) { currMV->x = EVEN(currMV->x); currMV->y = EVEN(currMV->y); } - - if (currMV->x > max_dx) - currMV->x=max_dx; - if (currMV->x < min_dx) - currMV->x=min_dx; - if (currMV->y > max_dy) - currMV->y=max_dy; - if (currMV->y < min_dy) - currMV->y=min_dy; -/***************** This is predictor SET A: only median prediction ******************/ + if (currMV->x > max_dx) + currMV->x = max_dx; + if (currMV->x < min_dx) + currMV->x = min_dx; + if (currMV->y > max_dy) + currMV->y = max_dy; + if (currMV->y < min_dy) + currMV->y = min_dy; + +/***************** This is predictor SET A: only median prediction ******************/ + + + iMinSAD = + sad8(cur, + get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, + iEdgedWidth), iEdgedWidth); + iMinSAD += + calc_delta_8(currMV->x - center_x, currMV->y - center_y, + (uint8_t) iFcode, iQuant); - - iMinSAD = sad8( cur, - 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, (uint8_t)iFcode, iQuant); - // thresh1 is fixed to 256 - if (iMinSAD < 256/4 ) - { - if (MotionFlags & PMV_QUICKSTOP8) - goto EPZS8_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP8) - goto EPZS8_Terminate_with_Refine; - } + if (iMinSAD < 256 / 4) { + if (MotionFlags & PMV_QUICKSTOP8) + goto EPZS8_Terminate_without_Refine; + if (MotionFlags & PMV_EARLYSTOP8) + goto EPZS8_Terminate_with_Refine; + } -/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ +/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ // MV=(0,0) is often a good choice CHECK_MV8_ZERO; // previous frame MV - CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); - + CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, prevMB->mvs[iSubBlock].y); + // left neighbour, if allowed - if (psad[1] != MV_MAX_ERROR) - { - if (!(MotionFlags & PMV_HALFPEL8 )) - { pmv[1].x = EVEN(pmv[1].x); + if (psad[1] != MV_MAX_ERROR) { + if (!(MotionFlags & PMV_HALFPEL8)) { + pmv[1].x = EVEN(pmv[1].x); pmv[1].y = EVEN(pmv[1].y); } - CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); + CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); } - // top neighbour, if allowed - if (psad[2] != MV_MAX_ERROR) - { - if (!(MotionFlags & PMV_HALFPEL8 )) - { pmv[2].x = EVEN(pmv[2].x); + if (psad[2] != MV_MAX_ERROR) { + if (!(MotionFlags & PMV_HALFPEL8)) { + pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); } - CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); - + CHECK_MV8_CANDIDATE(pmv[2].x, pmv[2].y); + // top right neighbour, if allowed - if (psad[3] != MV_MAX_ERROR) - { - if (!(MotionFlags & PMV_HALFPEL8 )) - { pmv[3].x = EVEN(pmv[3].x); + if (psad[3] != MV_MAX_ERROR) { + if (!(MotionFlags & PMV_HALFPEL8)) { + pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); } - CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); + CHECK_MV8_CANDIDATE(pmv[3].x, pmv[3].y); } } @@ -1894,78 +2726,75 @@ if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) iMinSAD -= MV8_00_BIAS; -*/ +*/ /* Terminate if MinSAD <= T_2 Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] */ - if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ - { - if (MotionFlags & PMV_QUICKSTOP8) - goto EPZS8_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP8) - goto EPZS8_Terminate_with_Refine; - } + if (iMinSAD < 512 / 4) { /* T_2 == 512/4 hardcoded */ + if (MotionFlags & PMV_QUICKSTOP8) + goto EPZS8_Terminate_without_Refine; + if (MotionFlags & PMV_EARLYSTOP8) + goto EPZS8_Terminate_with_Refine; + } /************ (Diamond Search) **************/ - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ if (!(MotionFlags & PMV_HALFPELDIAMOND8)) iDiamondSize *= 2; - + /* default: use best prediction as starting point for one call of EPZS_MainSearch */ -/* // there is no EPZS^2 for inter4v at the moment +// there is no EPZS^2 for inter4v at the moment + + if (MotionFlags & PMV_USESQUARES8) + MainSearchPtr = Square8_MainSearch; + else - if (MotionFlags & PMV_USESQUARES8) - EPZSMainSearchPtr = Square8_MainSearch; + if (MotionFlags & PMV_ADVANCEDDIAMOND8) + MainSearchPtr = AdvDiamond8_MainSearch; else -*/ + MainSearchPtr = Diamond8_MainSearch; - EPZSMainSearchPtr = Diamond8_MainSearch; - - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV->x, currMV->y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - iDiamondSize, iFcode, iQuant, 0); + 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, 0); - if (iSAD < iMinSAD) - { + if (iSAD < iMinSAD) { *currMV = newMV; iMinSAD = iSAD; } - if (MotionFlags & PMV_EXTSEARCH8) - { + if (MotionFlags & PMV_EXTSEARCH8) { /* extended mode: search (up to) two more times: orignal prediction and (0,0) */ - if (!(MVequal(pmv[0],backupMV)) ) - { - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); - - if (iSAD < iMinSAD) - { + 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, 0); + + if (iSAD < iMinSAD) { *currMV = newMV; iMinSAD = iSAD; } } - if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) - { - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - 0, 0, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); - - if (iSAD < iMinSAD) - { + 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, 0); + + if (iSAD < iMinSAD) { *currMV = newMV; iMinSAD = iSAD; } @@ -1974,139 +2803,647 @@ /*************** Choose best MV found **************/ -EPZS8_Terminate_with_Refine: - 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, iFcode, iQuant, iEdgedWidth); + EPZS8_Terminate_with_Refine: + if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step + iMinSAD = + Halfpel8_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); -EPZS8_Terminate_without_Refine: + 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 { + 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) ***************************************************************/ +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; -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; + const int32_t iWidth = pParam->width; + const int32_t iHeight = pParam->height; - int32_t f_sad16; - int32_t b_sad16; - int32_t i_sad16; - int32_t d_sad16; - int32_t best_sad; + 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 almost linear motion) */ + int best_sad; + + VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ + VECTOR f_interpolMV, b_interpolMV; VECTOR pmv_dontcare; + int min_dx, max_dx, min_dy, max_dy; + int f_min_dx, f_max_dx, f_min_dy, f_max_dy; + int b_min_dx, b_max_dx, b_min_dy, b_max_dy; + + int f_count=0; + int b_count=0; + int i_count=0; + int d_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++) - { - 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) - { + 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), cpb=0 without further ado */ + + 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; + mb->b_mvs[0] = mb->mvs[0] = zeroMV; continue; } + if (b_mb->mode == MODE_INTER4V) + { + d_sad16 = 0; + /* 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->deltamv.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, - frame->motion_flags, frame->quant, frame->fcode, - pParam, - f_mbs, f_mbs /* todo */, - &mb->mvs[0], &pmv_dontcare); // ignore pmv + &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, + &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); + + get_range(&f_min_dx, &f_max_dx, &f_min_dy, &f_max_dy, i, j, 16, iWidth, iHeight, + frame->fcode); + get_range(&b_min_dx, &b_max_dx, &b_min_dy, &b_max_dy, i, j, 16, iWidth, iHeight, + frame->bcode); + +/* Interpolated MC motion vector search, this is tedious and more complicated because there are + two values for everything, always one for backward and one for forward ME. Still, we don't gain + much from this search, maybe it should simply be skipped and simply current i_sad16 value used + as "optimal". */ + + i_sad16 = Diamond16_InterpolMainSearch( + f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + frame->image.y + i * 16 + j * 16 * edged_width, + b_ref->y, b_refH->y, b_refV->y, b_refHV->y, 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) - { + mb->mvs[0].x, mb->mvs[0].y, + mb->b_mvs[0].x, mb->b_mvs[0].y, + i_sad16, + &f_interpolMV, &b_interpolMV, + f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, + f_min_dx, f_max_dx, f_min_dy, f_max_dy, + b_min_dx, b_max_dx, b_min_dy, b_max_dy, + edged_width, 2, + frame->fcode, frame->bcode,frame->quant,0); + + i_sad16 = Diamond16_InterpolMainSearch( + f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + frame->image.y + i * 16 + j * 16 * edged_width, + b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + i, j, + f_interpolMV.x, f_interpolMV.y, + b_interpolMV.x, b_interpolMV.y, + i_sad16, + &f_interpolMV, &b_interpolMV, + f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, + f_min_dx, f_max_dx, f_min_dy, f_max_dy, + b_min_dx, b_max_dx, b_min_dy, b_max_dy, + edged_width, 1, + frame->fcode, frame->bcode,frame->quant,0); // equiv to halfpel refine + + +/* DIRECT MODE DELTA VECTOR SEARCH. + This has to be made more effective, but at the moment I'm happy it's running at all */ + +/* There are two range restrictions for direct mode: deltaMV is limited to [-32,31] in halfpel units, and + absolute vector must not lie outside of image dimensions. Constraint one is dealt with by CHECK_MV16_DIRECT + and for constraint two we need distance to boundary. This is done by get_range very large fcode (hack!) */ + + get_range(&min_dx, &max_dx, &min_dy, &max_dy, i, j, 16, iWidth, iHeight, 19); + + d_sad16 = Diamond16_DirectMainSearch( + f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + frame->image.y + i*16 + j*16*edged_width, + b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + i, j, + TRB,TRD, + 0,0, + d_sad16, + &mb->deltamv, + mb->directmv, // this has to be pre-initialized with b_mb->mvs[] + min_dx, max_dx, min_dy, max_dy, + edged_width, 2, frame->quant, 0); + + d_sad16 = Diamond16_DirectMainSearch( + f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + frame->image.y + i*16 + j*16*edged_width, + b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + i, j, + TRB,TRD, + mb->deltamv.x, mb->deltamv.y, + d_sad16, + &mb->deltamv, + mb->directmv, // this has to be pre-initialized with b_mb->mvs[] + min_dx, max_dx, min_dy, max_dy, + edged_width, 1, frame->quant, 0); // equiv to halfpel refine + + +// i_sad16 = 65535; /* remove the comment to disable any of the MODEs */ +// f_sad16 = 65535; +// b_sad16 = 65535; +// d_sad16 = 65535; + + if (f_sad16 < b_sad16) { best_sad = f_sad16; mb->mode = MODE_FORWARD; - } - else - { + } else { best_sad = b_sad16; mb->mode = MODE_BACKWARD; } - - if (i_sad16 < best_sad) - { + + if (i_sad16 < best_sad) { best_sad = i_sad16; mb->mode = MODE_INTERPOLATE; } - if (d_sad16 < best_sad) - { + if (d_sad16 < best_sad) { + + if (b_mb->mode == MODE_INTER4V) + { + + /* how to calc vectors is defined in standard. mvs[] and b_mvs[] are only for motion compensation */ + /* for the bitstream, the value mb->deltamv is read directly */ + + 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->deltamv.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->deltamv.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; } + 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++; + mb->mvs[0] = f_interpolMV; + mb->b_mvs[0] = b_interpolMV; + f_predMV = mb->mvs[0]; + b_predMV = mb->b_mvs[0]; + break; + case MODE_DIRECT: + d_count++; + break; + default: + break; + } + } } + +#ifdef _DEBUG_BFRAME_STAT + fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D: %04d\n", + f_count,b_count,i_count,d_count); +#endif + }