--- trunk/xvidcore/src/motion/motion_est.c 2002/04/11 15:04:05 117 +++ trunk/xvidcore/src/motion/motion_est.c 2002/04/25 06:55:00 136 @@ -2,6 +2,9 @@ * * Modifications: * + * 25.04.2002 partial prevMB conversion + * 22.04.2002 remove some compile warning by chenm001 + * 14.04.2002 added MotionEstimationBVOP() * 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between * EPZS and EPZS^2 * 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop @@ -40,6 +43,7 @@ #include "../prediction/mbprediction.h" #include "../global.h" #include "../utils/timer.h" +#include "motion.h" #include "sad.h" // very large value @@ -68,11 +72,6 @@ #define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) -#define MIN(X, Y) ((X)<(Y)?(X):(Y)) -#define MAX(X, Y) ((X)>(Y)?(X):(Y)) -#define ABS(X) (((X)>0)?(X):-(X)) -#define SIGN(X) (((X)>0)?1:-1) - int32_t PMVfastSearch16( const uint8_t * const pRef, const uint8_t * const pRefH, @@ -81,8 +80,11 @@ 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, - MACROBLOCK * const pMBs, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, VECTOR * const currMV, VECTOR * const currPMV); @@ -94,8 +96,11 @@ 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, - MACROBLOCK * const pMBs, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, VECTOR * const currMV, VECTOR * const currPMV); @@ -107,10 +112,13 @@ const uint8_t * const pRefHV, const IMAGE * const pCur, const int x, const int y, - const int start_x, int start_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, - MACROBLOCK * const pMBs, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, VECTOR * const currMV, VECTOR * const currPMV); @@ -121,10 +129,13 @@ const uint8_t * const pRefHV, const IMAGE * const pCur, const int x, const int y, - const int start_x, int start_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, - MACROBLOCK * const pMBs, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, VECTOR * const currMV, VECTOR * const currPMV); @@ -221,90 +232,6 @@ -/* calculate the min/max range (in halfpixels) - relative to the _MACROBLOCK_ position -*/ - -static void __inline get_range( - int32_t * const min_dx, int32_t * const max_dx, - int32_t * const min_dy, int32_t * const max_dy, - const uint32_t x, const uint32_t y, - const uint32_t block_sz, // block dimension, 8 or 16 - const uint32_t width, const uint32_t height, - const uint32_t fcode) -{ - - const int search_range = 32 << (fcode - 1); - const int high = search_range - 1; - const int low = -search_range; - - // convert full-pixel measurements to half pixel - const int hp_width = 2 * width; - const int hp_height = 2 * height; - const int hp_edge = 2 * block_sz; - const int hp_x = 2 * (x) * block_sz; // we need _right end_ of block, not x-coordinate - const int hp_y = 2 * (y) * block_sz; // same for _bottom end_ - - *max_dx = MIN(high, hp_width - hp_x); - *max_dy = MIN(high, hp_height - hp_y); - *min_dx = MAX(low, -(hp_edge + hp_x)); - *min_dy = MAX(low, -(hp_edge + hp_y)); - -} - - -/* - * getref: calculate reference image pointer - * the decision to use interpolation h/v/hv or the normal image is - * based on dx & dy. - */ - -static __inline const uint8_t * get_ref( - const uint8_t * const refn, - const uint8_t * const refh, - const uint8_t * const refv, - const uint8_t * const refhv, - const uint32_t x, const uint32_t y, - const uint32_t block, // block dimension, 8 or 16 - const int32_t dx, const int32_t dy, - const uint32_t stride) -{ - - switch ( ((dx&1)<<1) + (dy&1) ) // ((dx%2)?2:0)+((dy%2)?1:0) - { - case 0 : return refn + (x*block+dx/2) + (y*block+dy/2)*stride; - case 1 : return refv + (x*block+dx/2) + (y*block+(dy-1)/2)*stride; - case 2 : return refh + (x*block+(dx-1)/2) + (y*block+dy/2)*stride; - default : - case 3 : return refhv + (x*block+(dx-1)/2) + (y*block+(dy-1)/2)*stride; - } - -} - - -/* This is somehow a copy of get_ref, but with MV instead of X,Y */ - -static __inline const uint8_t * get_ref_mv( - const uint8_t * const refn, - const uint8_t * const refh, - const uint8_t * const refv, - const uint8_t * const refhv, - const uint32_t x, const uint32_t y, - const uint32_t block, // block dimension, 8 or 16 - const VECTOR* mv, // measured in half-pel! - const uint32_t stride) -{ - - switch ( (((mv->x)&1)<<1) + ((mv->y)&1) ) - { - case 0 : return refn + (x*block+(mv->x)/2) + (y*block+(mv->y)/2)*stride; - case 1 : return refv + (x*block+(mv->x)/2) + (y*block+((mv->y)-1)/2)*stride; - case 2 : return refh + (x*block+((mv->x)-1)/2) + (y*block+(mv->y)/2)*stride; - default : - case 3 : return refhv + (x*block+((mv->x)-1)/2) + (y*block+((mv->y)-1)/2)*stride; - } - -} #ifndef SEARCH16 #define SEARCH16 PMVfastSearch16 @@ -318,18 +245,23 @@ #endif bool MotionEstimation( - MACROBLOCK * const pMBs, MBParam * const pParam, - const IMAGE * const pRef, + FRAMEINFO * const current, + FRAMEINFO * const reference, const IMAGE * const pRefH, const IMAGE * const pRefV, const IMAGE * const pRefHV, - IMAGE * const pCurrent, 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; @@ -342,16 +274,49 @@ if (sadInit) (*sadInit)(); + + + /* eventhough we have a seperate prevMBs, + pmvfast/epsz does something "funny" with the previous frames data */ + + for (i = 0; i < iHcount; i++) + for (j = 0; j < iWcount; j++) + { + pMBs[j + i * iWcount].mvs[0] = prevMBs[j + i * iWcount].mvs[0]; + pMBs[j + i * iWcount].mvs[1] = prevMBs[j + i * iWcount].mvs[1]; + pMBs[j + i * iWcount].mvs[2] = prevMBs[j + i * iWcount].mvs[2]; + pMBs[j + i * iWcount].mvs[3] = prevMBs[j + i * iWcount].mvs[3]; + } + + /*dprintf("*** BEFORE ***"); + for (i = 0; i < iHcount; i++) + for (j = 0; j < iWcount; j++) + { + dprintf(" [%i,%i] mode=%i dquant=%i mvs=(%i %i %i %i) sad8=(%i %i %i %i) sad16=(%i)", j,i, + pMBs[j + i * iWcount].mode, + pMBs[j + i * iWcount].dquant, + pMBs[j + i * iWcount].mvs[0], + pMBs[j + i * iWcount].mvs[1], + pMBs[j + i * iWcount].mvs[2], + pMBs[j + i * iWcount].mvs[3], + prevMBs[j + i * iWcount].sad8[0], + prevMBs[j + i * iWcount].sad8[1], + prevMBs[j + i * iWcount].sad8[2], + prevMBs[j + i * iWcount].sad8[3], + prevMBs[j + i * iWcount].sad16); + } + */ // 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]; sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - j, i, pParam->motion_flags, - pParam, pMBs, &mv16, &pmv16); + j, i, current->motion_flags, current->quant, current->fcode, + pParam, pMBs, prevMBs, &mv16, &pmv16); pMB->sad16=sad16; @@ -367,6 +332,8 @@ 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; @@ -374,23 +341,27 @@ continue; } - if (pParam->global_flags & XVID_INTER4V) + 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, pParam->motion_flags, - pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); + 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, pParam->motion_flags, - pParam, pMBs, &pMB->mvs[1], &pMB->pmvs[1]); + 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, pParam->motion_flags, - pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); + 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, pParam->motion_flags, - pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); + 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]); sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; } @@ -400,9 +371,11 @@ mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v */ - if (pMB->dquant == NO_CHANGE) { - if (((pParam->global_flags & XVID_INTER4V)==0) || - (sad16 < (sad8 + (int32_t)(IMV16X16 * pParam->quant)))) { + 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; @@ -425,6 +398,25 @@ } } +/* dprintf("*** AFTER ***", pMBs[0].b_mvs[0].x); + for (i = 0; i < iHcount; i++) + for (j = 0; j < iWcount; j++) + { + dprintf(" [%i,%i] mode=%i dquant=%i mvs=(%i %i %i %i) sad8=(%i %i %i %i) sad16=(%i)", j,i, + pMBs[j + i * iWcount].mode, + pMBs[j + i * iWcount].dquant, + pMBs[j + i * iWcount].mvs[0], + pMBs[j + i * iWcount].mvs[1], + pMBs[j + i * iWcount].mvs[2], + pMBs[j + i * iWcount].mvs[3], + pMBs[j + i * iWcount].sad8[0], + pMBs[j + i * iWcount].sad8[1], + pMBs[j + i * iWcount].sad8[2], + pMBs[j + i * iWcount].sad8[3], + pMBs[j + i * iWcount].sad16); + } + */ + return 0; } @@ -538,13 +530,15 @@ const IMAGE * const pCur, const int x, const int y, const uint32_t MotionFlags, + const uint32_t iQuant, + const uint32_t iFcode, MBParam * const pParam, - MACROBLOCK * const pMBs, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, VECTOR * const currMV, VECTOR * const currPMV) { const int32_t iEdgedWidth = pParam->edged_width; - const int32_t iQuant = pParam->quant; const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; int32_t iSAD; int32_t pred_x,pred_y; @@ -866,14 +860,15 @@ 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, - MACROBLOCK * const pMBs, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, VECTOR * const currMV, VECTOR * const currPMV) { - const uint32_t iWcount = pParam->mb_width; - const int32_t iFcode = pParam->fixed_code; - const int32_t iQuant = pParam->quant; + 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; @@ -895,7 +890,8 @@ VECTOR pmv[4]; int32_t psad[4]; - MACROBLOCK * const pMB = pMBs + x + y * iWcount; + const MACROBLOCK * const pMB = pMBs + x + y * iWcount; + const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; static int32_t threshA,threshB; int32_t bPredEq; @@ -939,7 +935,7 @@ If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 */ - if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[0]) ) ) + 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. @@ -993,7 +989,7 @@ 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,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) + if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) { if (MotionFlags & PMV_QUICKSTOP16) @@ -1014,7 +1010,7 @@ CHECK_MV16_ZERO; // previous frame MV is always possible - CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); + CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); // left neighbour, if allowed if (x != 0) @@ -1036,7 +1032,7 @@ CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); // top right neighbour, if allowed - if (x != (iWcount-1)) + if ((uint32_t)x != (iWcount-1)) { if (!(MotionFlags & PMV_HALFPEL16 )) { pmv[3].x = EVEN(pmv[3].x); @@ -1050,7 +1046,7 @@ If Motion Vector equal to Previous frame motion vector and MinSADmvs[0]) && (iMinSAD < pMB->sad16) ) ) + if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast16_Terminate_without_Refine; @@ -1236,17 +1232,17 @@ const uint8_t * const pRefHV, const IMAGE * const pCur, const int x, const int y, - const int start_x, int start_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, - MACROBLOCK * const pMBs, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, VECTOR * const currMV, VECTOR * const currPMV) { - const uint32_t iWcount = pParam->mb_width; - - const int32_t iFcode = pParam->fixed_code; - const int32_t iQuant = pParam->quant; + 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; @@ -1265,7 +1261,8 @@ VECTOR newMV; VECTOR backupMV; - MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; + const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; + const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; static int32_t threshA,threshB; int32_t iFound,bPredEq; @@ -1344,7 +1341,7 @@ 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,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) + if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; @@ -1403,7 +1400,7 @@ If Motion Vector equal to Previous frame motion vector and MinSADmvs[iSubBlock]) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) + if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; @@ -1493,15 +1490,16 @@ 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, - MACROBLOCK * const pMBs, + 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 int32_t iFcode = pParam->fixed_code; - const int32_t iQuant = pParam->quant; + 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; @@ -1521,7 +1519,8 @@ int32_t psad[8]; static MACROBLOCK * oldMBs = NULL; - MACROBLOCK * const pMB = pMBs + x + y * iWcount; + const MACROBLOCK * const pMB = pMBs + x + y * iWcount; + const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; MACROBLOCK * oldMB = NULL; static int32_t thresh2; @@ -1584,7 +1583,7 @@ iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; // thresh1 is fixed to 256 - if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) + if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) { if (MotionFlags & PMV_QUICKSTOP16) goto EPZS16_Terminate_without_Refine; @@ -1636,7 +1635,7 @@ CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); // top right neighbour, if allowed - if (x != (iWcount-1)) + if ((uint32_t)x != (iWcount-1)) { if (!(MotionFlags & PMV_HALFPEL16 )) { pmv[3].x = EVEN(pmv[3].x); @@ -1651,7 +1650,7 @@ */ if ( (iMinSAD <= thresh2) - || ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD <= pMB->sad16) ) ) + || ( MVequal(*currMV,pMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) { if (MotionFlags & PMV_QUICKSTOP16) goto EPZS16_Terminate_without_Refine; @@ -1677,11 +1676,11 @@ // right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs - if (x != iWcount-1) + if ((uint32_t)x != iWcount-1) CHECK_MV16_CANDIDATE((pMB+1)->mvs[0].x,oldMB->mvs[0].y); // bottom neighbour, dito - if (y != iHcount-1) + if ((uint32_t)y != iHcount-1) CHECK_MV16_CANDIDATE((pMB+iWcount)->mvs[0].x,oldMB->mvs[0].y); /* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ @@ -1777,15 +1776,15 @@ 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, - MACROBLOCK * const pMBs, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, VECTOR * const currMV, VECTOR * const currPMV) { - const uint32_t iWcount = pParam->mb_width; - const int32_t iFcode = pParam->fixed_code; - const int32_t iQuant = pParam->quant; - + 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; @@ -1807,7 +1806,8 @@ const int32_t iSubBlock = ((y&1)<<1) + (x&1); - MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; + 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; @@ -1970,3 +1970,127 @@ return iMinSAD; } + + + + +/* *********************************************************** + bvop motion estimation +// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) +***************************************************************/ + +/* +void MotionEstimationBVOP( + MBParam * const pParam, + FRAMEINFO * const frame, + + // forward (past) reference + const MACROBLOCK * const f_mbs, + const IMAGE * const f_ref, + const IMAGE * const f_refH, + const IMAGE * const f_refV, + const IMAGE * const f_refHV, + // backward (future) reference + const MACROBLOCK * const b_mbs, + const IMAGE * const b_ref, + const IMAGE * const b_refH, + const IMAGE * const b_refV, + const IMAGE * const b_refHV) +{ + const uint32_t mb_width = pParam->mb_width; + const uint32_t mb_height = pParam->mb_height; + const int32_t edged_width = pParam->edged_width; + + int32_t i,j; + + int32_t f_sad16; + int32_t b_sad16; + int32_t i_sad16; + int32_t d_sad16; + int32_t best_sad; + + VECTOR pmv_dontcare; + + // note: i==horizontal, j==vertical + for (j = 0; j < mb_height; j++) + { + for (i = 0; i < mb_width; i++) + { + MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; + const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; + const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; + + if (b_mb->mode == MODE_INTER + && b_mb->cbp == 0 + && b_mb->mvs[0].x == 0 + && b_mb->mvs[0].y == 0) + { + mb->mode = MB_IGNORE; + mb->mvs[0].x = 0; + mb->mvs[0].y = 0; + mb->b_mvs[0].x = 0; + mb->b_mvs[0].y = 0; + continue; + } + + + // forward search + f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + &frame->image, + i, j, + frame->motion_flags, frame->quant, frame->fcode, + pParam, + f_mbs, + &mb->mvs[0], &pmv_dontcare); // ignore pmv + + // backward search + b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + &frame->image, + i, j, + frame->motion_flags, frame->quant, frame->bcode, + pParam, + b_mbs, + &mb->b_mvs[0], &pmv_dontcare); // ignore pmv + + // interpolate search (simple, but effective) + i_sad16 = sad16bi_c( + frame->image.y + i*16 + j*16*edged_width, + get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + i, j, 16, mb->mvs[0].x, mb->mvs[0].y, edged_width), + get_ref(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + i, j, 16, mb->b_mvs[0].x, mb->b_mvs[0].x, edged_width), + edged_width); + + // TODO: direct search + // predictor + range of [-32,32] + d_sad16 = 65535; + + + if (f_sad16 < b_sad16) + { + best_sad = f_sad16; + mb->mode = MB_FORWARD; + } + else + { + best_sad = b_sad16; + mb->mode = MB_BACKWARD; + } + + if (i_sad16 < best_sad) + { + best_sad = i_sad16; + mb->mode = MB_INTERPOLATE; + } + + if (d_sad16 < best_sad) + { + best_sad = d_sad16; + mb->mode = MB_DIRECT; + } + + } + } +} + +*/