--- branches/dev-api-3/xvidcore/src/motion/motion_est.c	2002/09/28 17:28:18	568
+++ branches/dev-api-3/xvidcore/src/motion/motion_est.c	2002/11/07 10:31:03	631
@@ -37,6 +37,7 @@
 #include "../prediction/mbprediction.h"
 #include "../global.h"
 #include "../utils/timer.h"
+#include "../image/interpolate8x8.h"
 #include "motion_est.h"
 #include "motion.h"
 #include "sad.h"
@@ -51,6 +52,16 @@
 #define CHECK_CANDIDATE(X,Y,D) { \
 (*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); }
 
+#define GET_REFERENCE(X, Y, REF) { \
+	switch ( (((X)&1)<<1) + ((Y)&1) ) \
+	{ \
+		case 0 : REF = (uint8_t *)data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \
+		case 1 : REF = (uint8_t *)data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \
+		case 2 : REF = (uint8_t *)data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \
+		default : REF = (uint8_t *)data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \
+	} \
+}
+
 #define iDiamondSize 2
 
 static __inline int
@@ -78,13 +89,50 @@
 	return xb + yb;
 }
 
+static int32_t 
+ChromaSAD(int dx, int dy, const SearchData * const data)
+{
+	int sad;
+	dx = (dx >> 1) + roundtab_79[dx & 0x3];
+	dy = (dy >> 1) + roundtab_79[dy & 0x3];
+
+	switch (((dx & 1) << 1) + (dy & 1))	{ // ((dx%2)?2:0)+((dy%2)?1:0)
+		case 0:
+			sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2);
+			sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2);
+			break;
+		case 1:
+			dx = dx / 2; dy = (dy - 1) / 2;
+			sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2);
+			sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2);
+			break;
+		case 2:
+			dx = (dx - 1) / 2; dy = dy / 2;
+			sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2);
+			sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2);
+			break;
+		default:
+			dx = (dx - 1) / 2; dy = (dy - 1) / 2;
+			interpolate8x8_halfpel_hv(data->RefQ,
+									 data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2,
+									 data->rounding);
+			sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2);
+			interpolate8x8_halfpel_hv(data->RefQ,
+									 data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2,
+									 data->rounding);
+			sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2);
+			break;
+	}
+	return sad;
+}
+
 
 /* CHECK_CANDIATE FUNCTIONS START */
 
+
 static void 
 CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
 {
-	int32_t * const sad = data->temp;
 	int t;
 	const uint8_t * Reference;
 
@@ -98,11 +146,15 @@
 		default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break;
 	}
 	
-	data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1);
+	data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
+
+	if (data->qpel) t = d_mv_bits(2*x - data->predQMV.x, 2*y - data->predQMV.y, data->iFcode);
+	else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode);
 
-	t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode);
-	data->temp[0] += lambda_vec16[data->iQuant] * t;
-	data->temp[1] += lambda_vec8[data->iQuant] * t;
+	data->temp[0] += (data->lambda16 * t * data->temp[0])/1000;
+	data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100;
+
+	if (data->chroma) data->temp[0] += ChromaSAD(x, y, data);
 
 	if (data->temp[0] < data->iMinSAD[0]) {
 		data->iMinSAD[0] = data->temp[0];
@@ -137,9 +189,8 @@
 		default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break;
 	}
 
-	sad = lambda_vec16[data->iQuant] *
-			d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode);
-	sad += sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
+	sad = sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR);
+	sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000;
 
 	if (sad < *(data->iMinSAD)) {
 		*(data->iMinSAD) = sad;
@@ -148,6 +199,84 @@
 }
 
 static void 
+CheckCandidate16_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
+
+// CheckCandidate16 variant which expects x and y in quarter pixel resolution
+// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!)
+// around currentMV!
+{
+	int t;
+	uint8_t * Reference = (uint8_t *)data->RefQ;
+	const uint8_t *ref1, *ref2, *ref3, *ref4;
+	VECTOR halfpelMV = *(data->currentMV);
+
+	int32_t iEdgedWidth = data->iEdgedWidth;
+	uint32_t rounding = data->rounding;
+
+	if (( x > data->max_dx) || ( x < data->min_dx)
+		|| ( y > data->max_dy) || (y < data->min_dy)) return;
+
+	GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases
+	switch( ((x&1)<<1) + (y&1) )
+	{
+	case 0: // pure halfpel position - shouldn't happen during a refinement step
+		GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); 
+		break;
+
+	case 1: // x halfpel, y qpel - top or bottom during qpel refinement
+		GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
+		interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding);
+		interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding);
+		interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding);
+		interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding);
+		break;
+
+	case 2: // x qpel, y halfpel - left or right during qpel refinement
+		GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2);
+		interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding);
+		interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding);
+		interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding);
+		interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding);
+		break;
+
+	default: // x and y in qpel resolution - the "corners" (top left/right and
+			 // bottom left/right) during qpel refinement
+		GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
+		GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3);
+		GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4);
+
+		interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
+		interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
+		interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
+		interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
+		break;
+	}
+	
+	data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp+1);
+
+	t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode);
+	data->temp[0] += (data->lambda16 * t * data->temp[0])/1000;
+	data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100;
+
+	if (data->chroma)
+		data->temp[0] += ChromaSAD(x/2, y/2, data);
+
+	if (data->temp[0] < data->iMinSAD[0]) {
+		data->iMinSAD[0] = data->temp[0];
+		data->currentQMV[0].x = x; data->currentQMV[0].y = y;
+	/*	*dir = Direction;*/ }
+
+	if (data->temp[1] < data->iMinSAD[1]) {
+		data->iMinSAD[1] = data->temp[1]; data->currentQMV[1].x = x; data->currentQMV[1].y = y; }
+	if (data->temp[2] < data->iMinSAD[2]) {
+		data->iMinSAD[2] = data->temp[2]; data->currentQMV[2].x = x; data->currentQMV[2].y = y; }
+	if (data->temp[3] < data->iMinSAD[3]) {
+		data->iMinSAD[3] = data->temp[3]; data->currentQMV[3].x = x; data->currentQMV[3].y = y; }
+	if (data->temp[4] < data->iMinSAD[4]) {
+		data->iMinSAD[4] = data->temp[4]; data->currentQMV[4].x = x; data->currentQMV[4].y = y; }
+}
+
+static void 
 CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
 {
 	int32_t sad;
@@ -155,10 +284,7 @@
 	if (( x > data->max_dx) || ( x < data->min_dx)
 		|| ( y > data->max_dy) || (y < data->min_dy)) return;
 
-	sad = lambda_vec16[data->iQuant] *
-			d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode);
-
-	sad += sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
+	sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
 					data->iEdgedWidth, 256*4096);
 
 	if (sad < *(data->iMinSAD)) {
@@ -193,11 +319,11 @@
 		default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break;
 	}
 
-	sad = lambda_vec16[data->iQuant] *
-			( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + 
-			  d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) );
+	sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
 
-	sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
+	sad += (data->lambda16 *
+			( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + 
+			  d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000;
 
 	if (sad < *(data->iMinSAD)) {
 		*(data->iMinSAD) = sad;
@@ -208,7 +334,7 @@
 static void 
 CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
 {
-	int32_t sad;
+	int32_t sad = 0;
 	int k;
 	const uint8_t *ReferenceF;
 	const uint8_t *ReferenceB;
@@ -216,8 +342,6 @@
 
 	if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
 
-	sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1);
-
 	for (k = 0; k < 4; k++) {
 		mvs.x = data->directmvF[k].x + x;
 		b_mvs.x = ((x == 0) ?
@@ -255,6 +379,8 @@
 		if (sad > *(data->iMinSAD)) return;
 	}
 
+	sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000;
+
 	if (sad < *(data->iMinSAD)) {
 		*(data->iMinSAD) = sad;
 		data->currentMV->x = x; data->currentMV->y = y;
@@ -269,9 +395,7 @@
 	const uint8_t *ReferenceB;
 	VECTOR mvs, b_mvs;
 
-	if (( x > 31) || ( x < -31) || ( y > 31) || (y < -31)) return;
-	
-		sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1);
+	if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
 
 	mvs.x = data->directmvF[0].x + x;
 	b_mvs.x = ((x == 0) ?
@@ -302,7 +426,8 @@
 		default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break;
 	}
 	
-	sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
+	sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
+	sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000;
 
 	if (sad < *(data->iMinSAD)) {
 		*(data->iMinSAD) = sad;
@@ -313,7 +438,7 @@
 static void
 CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
 {
-	int32_t sad;
+	int32_t sad; int t;
 	const uint8_t * Reference;
 
 	if (( x > data->max_dx) || ( x < data->min_dx)
@@ -328,7 +453,10 @@
 	}
 
 	sad = sad8(data->Cur, Reference, data->iEdgedWidth);
-	sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode);
+	if (data->qpel) t = d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode);
+	else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode);
+
+	sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100;
 
 	if (sad < *(data->iMinSAD)) {
 		*(data->iMinSAD) = sad;
@@ -336,6 +464,62 @@
 		*dir = Direction; }
 }
 
+static void
+CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
+// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution
+// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!)
+// around currentMV!
+
+{
+	int32_t sad;
+	uint8_t *Reference = (uint8_t *) data->RefQ;
+	const uint8_t *ref1, *ref2, *ref3, *ref4;
+	VECTOR halfpelMV = *(data->currentMV);
+
+	int32_t iEdgedWidth = data->iEdgedWidth;
+	uint32_t rounding = data->rounding;
+
+	if (( x > data->max_dx) || ( x < data->min_dx)
+		|| ( y > data->max_dy) || (y < data->min_dy)) return;
+
+	GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1);
+	switch( ((x&1)<<1) + (y&1) )
+	{
+	case 0: // pure halfpel position - shouldn't happen during a refinement step
+		GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); 
+		break;
+
+	case 1: // x halfpel, y qpel - top or bottom during qpel refinement
+		GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
+
+		interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding);
+		break;
+
+	case 2: // x qpel, y halfpel - left or right during qpel refinement
+		GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2);
+
+		interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding);
+		break;
+
+	default: // x and y in qpel resolution - the "corners" (top left/right and
+			 // bottom left/right) during qpel refinement
+		GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
+		GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3);
+		GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4);
+
+		interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
+		break;
+	}
+
+	sad = sad8(data->Cur, Reference, data->iEdgedWidth);
+	sad += (data->lambda8 * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode) * (sad+NEIGH_8X8_BIAS))/100;
+
+	if (sad < *(data->iMinSAD)) {
+		*(data->iMinSAD) = sad;
+		data->currentQMV->x = x; data->currentQMV->y = y;
+		*dir = Direction; }
+}
+
 /* CHECK_CANDIATE FUNCTIONS END */
 
 /* MAINSEARCH FUNCTIONS START */
@@ -504,6 +688,28 @@
 	CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0);
 }
 
+
+static void
+QuarterpelRefine(const SearchData * const data)
+{
+/* Perform quarter pixel refinement*/
+
+	VECTOR backupMV = *(data->currentQMV);
+	int iDirection; //not needed
+
+	CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0);
+	CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0);
+	CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0);
+	CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0);
+
+	CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0);
+	CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0);
+
+	CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0);
+	CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0);
+
+}
+
 static __inline int
 SkipDecisionP(const IMAGE * current, const IMAGE * reference,
 							const int x, const int y,
@@ -530,6 +736,10 @@
 	pMB->mode = MODE_NOT_CODED;
 	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->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0;
+	pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0;
+
 	pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
 }
 
@@ -550,62 +760,98 @@
 
 	uint32_t x, y;
 	uint32_t iIntra = 0;
-	int32_t InterBias;
+	int32_t InterBias, quant = current->quant, sad00;
+	uint8_t *qimage;
 
 	// some pre-initialized thingies for SearchP
 	int32_t temp[5];
 	VECTOR currentMV[5];
+	VECTOR currentQMV[5];
 	int32_t iMinSAD[5];
 	SearchData Data;
 	Data.iEdgedWidth = pParam->edged_width;
 	Data.currentMV = currentMV;
+	Data.currentQMV = currentQMV;
 	Data.iMinSAD = iMinSAD;
 	Data.temp = temp;
 	Data.iFcode = current->fcode;
-
+	Data.rounding = pParam->m_rounding_type;
+	Data.qpel = pParam->m_quarterpel;
+	Data.chroma = current->global_flags & XVID_ME_COLOUR;
+
+	if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL)
+		return 1; // allocate some mem for qpel interpolated blocks
+				  // somehow this is dirty since I think we shouldn't use malloc outside
+				  // encoder_create() - so please fix me!
+	Data.RefQ = qimage;
 	if (sadInit) (*sadInit) ();
 
 	for (y = 0; y < pParam->mb_height; y++)	{
 		for (x = 0; x < pParam->mb_width; x++)	{
-
 			MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
-			int32_t sad00 =  pMB->sad16 
+			
+			pMB->sad16 
 				= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16,
 							pRef->y + (x + y * pParam->edged_width) * 16,
 							pParam->edged_width, pMB->sad8 );
 
+			if (Data.chroma) {
+				pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8,
+								pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2);
+
+				pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8,
+								pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2);
+			}
+
+			sad00 = pMB->sad16; //if no gmc; else sad00 = (..)
+
 			if (!(current->global_flags & XVID_LUMIMASKING)) {
 				pMB->dquant = NO_CHANGE;
-				pMB->quant = current->quant; }
+				pMB->quant = current->quant; 
+			} else {
+				if (pMB->dquant != NO_CHANGE) {
+					quant += DQtab[pMB->dquant];
+					if (quant > 31) quant = 31;
+					else if (quant < 1) quant = 1;
+				}
+				pMB->quant = quant;
+			}
 
 //initial skip decision
-
-			if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant)
-				&& (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) {
-				if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) {
+/* no early skip for GMC (global vector = skip vector is unknown!)  */
+			if (current->coding_type == P_VOP)	{ /* no fast SKIP for S(GMC)-VOPs */
+				if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH)
+					if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) {
 						SkipMacroblockP(pMB, sad00);
 						continue;
-				}
-			} else sad00 = 256*4096; // skip not allowed - for final skip decision
-
-			SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
+					}
+			}
+			
+			SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
 						y, current->motion_flags, pMB->quant,
 						&Data, pParam, pMBs, reference->mbs,
 						current->global_flags & XVID_INTER4V, pMB);
 
 /* final skip decision, a.k.a. "the vector you found, really that good?" */
-			if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP)
-				if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH)
-				{ SkipMacroblockP(pMB, sad00); continue; }
+			if (current->coding_type == P_VOP)	{ 
+				if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP)
+				&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) )
+					if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) {
+						SkipMacroblockP(pMB, sad00);
+						continue;
+					}
+			}
 	
 /* finally, intra decision */
 
 			InterBias = MV16_INTER_BIAS;
-			if (pMB->quant > 8)  InterBias += 50 * (pMB->quant - 8); // to make high quants work
+			if (pMB->quant > 8)  InterBias += 80 * (pMB->quant - 8); // to make high quants work
 			if (y != 0)
-				if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50;
+				if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 80;
 			if (x != 0)
-				if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50;
+				if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 80;
+			
+			if (Data.chroma) InterBias += 50; // to compensate bigger SAD
 
 			if (InterBias < pMB->sad16)  {
 				const int32_t deviation =
@@ -613,16 +859,25 @@
 						  pParam->edged_width);
 
 				if (deviation < (pMB->sad16 - InterBias)) {
-					if (++iIntra >= iLimit) return 1;
+					if (++iIntra >= iLimit) { free(qimage); return 1; }
 					pMB->mode = MODE_INTRA;
 					pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] =
 							pMB->mvs[3] = zeroMV;
+					pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] =
+							pMB->qmvs[3] = zeroMV;
 					pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] =
 						pMB->sad8[3] = 0;
 				}
 			}
 		}
 	}
+	free(qimage);
+
+	if (current->coding_type == S_VOP)	/* first GMC step only for S(GMC)-VOPs */
+		current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode );
+	else
+		current->GMC_MV = zeroMV;
+
 	return 0;
 }
 
@@ -680,7 +935,7 @@
 }
 
 static void
-SearchP(const uint8_t * const pRef,
+SearchP(const IMAGE * const pRef,
 		const uint8_t * const pRefH,
 		const uint8_t * const pRefV,
 		const uint8_t * const pRefHV,
@@ -702,16 +957,23 @@
 
 	get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);  //has to be changed to get_pmv(2)()
 	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
-				pParam->width, pParam->height, Data->iFcode);
+				pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel);
 
 	Data->predMV = pmv[0];
+
 	Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16;
-	Data->Ref = pRef + (x + Data->iEdgedWidth*y)*16;
+	Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8;
+	Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8;
+
+	Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16;
 	Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16;
 	Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16;
 	Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16;
+	Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
+	Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
 
-	Data->iQuant = iQuant;
+	Data->lambda16 = lambda_vec16[iQuant];
+	Data->lambda8 = lambda_vec8[iQuant];
 
 	if (!(MotionFlags & PMV_HALFPEL16)) {
 		Data->min_dx = EVEN(Data->min_dx);
@@ -721,14 +983,16 @@
 	
 	if (pMB->dquant != NO_CHANGE) inter4v = 0;
 
-	if (inter4v) CheckCandidate = CheckCandidate16;
-	else CheckCandidate = CheckCandidate16no4v;
+	for(i = 0;  i < 5; i++)
+		Data->currentMV[i].x = Data->currentMV[i].y = 0;
 
-	for(i = 0;  i < 5; i++) Data->currentMV[i].x = Data->currentMV[i].y = 0;
+	if (pParam->m_quarterpel) {
+		Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
+		i = d_mv_bits(Data->predQMV.x, Data->predQMV.y, Data->iFcode);
+	} else i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode);
 
-	i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode);
-	Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i;
-	Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i;
+	Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000;
+	Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100;
 	Data->iMinSAD[2] = pMB->sad8[1];
 	Data->iMinSAD[3] = pMB->sad8[2];
 	Data->iMinSAD[4] = pMB->sad8[3];
@@ -742,10 +1006,9 @@
 	PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
 					prevMBs + x + y * pParam->mb_width);
 
-	if (inter4v) CheckCandidate = CheckCandidate16;
+	if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16;
 	else CheckCandidate = CheckCandidate16no4v;
 
-
 /* main loop. checking all predictions */
 
 	for (i = 1; i < 7; i++) {
@@ -779,7 +1042,7 @@
 			if (!(MVequal(startMV, backupMV))) {
 				bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
 
-				CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data);
+				(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data);
 				(*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
 				if (bSAD < Data->iMinSAD[0]) {
 					Data->currentMV[0] = backupMV;
@@ -792,7 +1055,7 @@
 			if (!(MVequal(startMV, backupMV))) {
 				bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
 
-				CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data);
+				(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data);
 				(*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
 				if (bSAD < Data->iMinSAD[0]) {
 					Data->currentMV[0] = backupMV;
@@ -803,15 +1066,48 @@
 
 	if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data);
 
+	for(i = 0; i < 5; i++) {
+		Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
+		Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
+	}
+
+	if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) {
+
+		CheckCandidate = CheckCandidate16_qpel;
+		get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
+				pParam->width, pParam->height, Data->iFcode, 0);
+
+		QuarterpelRefine(Data);
+	}
+
+	if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0;
 	if (inter4v) {
 		SearchData Data8;
 		Data8.iFcode = Data->iFcode;
-		Data8.iQuant = Data->iQuant;
+		Data8.lambda8 = Data->lambda8;
 		Data8.iEdgedWidth = Data->iEdgedWidth;
+		Data8.RefQ = Data->RefQ;
+		Data8.qpel = Data->qpel;
 		Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
 		Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
 		Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
 		Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
+		
+		if (Data->chroma) {
+			int sum, dx, dy;
+
+			if(pParam->m_quarterpel) {
+				sum = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2;
+			} else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y;
+			dy = (sum >> 3) + roundtab_76[sum & 0xf];
+
+			if(pParam->m_quarterpel) {
+				sum = pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2;
+			} else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x;
+			dx = (sum >> 3) + roundtab_76[sum & 0xf];
+			
+			Data->iMinSAD[1] += ChromaSAD(dx, dy, Data);
+		}
 	}
 
 	if (!(inter4v) ||
@@ -822,18 +1118,25 @@
 		pMB->mvs[0] = pMB->mvs[1]
 			= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
 
+		pMB->qmvs[0] = pMB->qmvs[1]
+			= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
+
 		pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] =
 			pMB->sad8[2] = pMB->sad8[3] =  Data->iMinSAD[0];
 
-		pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
-		pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
+		if(pParam->m_quarterpel) {
+			pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x;
+			pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y;
+		} else {
+			pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
+			pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
+		}
 	} else {
 // INTER4V MODE; all other things are already set in Search8 
 		pMB->mode = MODE_INTER4V;
 		pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + 
 			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant;
 	}
-
 }
 
 static void
@@ -846,15 +1149,23 @@
 		const int block,
 		SearchData * const Data)
 {
-	Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block);
 	Data->iMinSAD = OldData->iMinSAD + 1 + block;
 	Data->currentMV = OldData->currentMV + 1 + block;
+	Data->currentQMV = OldData->currentQMV + 1 + block;
 
-	if (block != 0)
-		*(Data->iMinSAD) += lambda_vec8[Data->iQuant] *
-								d_mv_bits(	Data->currentMV->x - Data->predMV.x, 
-											Data->currentMV->y - Data->predMV.y,
-											Data->iFcode);
+	if(pParam->m_quarterpel) {
+		Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block);
+		if (block != 0)	*(Data->iMinSAD) += (Data->lambda8 *
+									d_mv_bits(	Data->currentQMV->x - Data->predQMV.x, 
+												Data->currentQMV->y - Data->predQMV.y,
+												Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100;
+	} else {
+		Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block);
+		if (block != 0)	*(Data->iMinSAD) += (Data->lambda8 *
+									d_mv_bits(	Data->currentMV->x - Data->predMV.x, 
+												Data->currentMV->y - Data->predMV.y,
+												Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100;
+	}
 
 	if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) {
 
@@ -866,27 +1177,60 @@
 		Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1));
 		
 		get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
-				pParam->width, pParam->height, OldData->iFcode);
-
+				pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel);
 		CheckCandidate = CheckCandidate8;
 
 		if (MotionFlags & PMV_EXTSEARCH8) {
-
+			int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
+			
 			MainSearchFunc *MainSearchPtr;
 			if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
 				else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
 					else MainSearchPtr = DiamondSearch;
 
-			(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255);	}
+			(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255);
+
+			if(*(Data->iMinSAD) < temp_sad) {
+					Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
+					Data->currentQMV->y = 2 * Data->currentMV->y;
+			}
+		}
+
+		if (MotionFlags & PMV_HALFPELREFINE8) {
+			int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
+
+			HalfpelRefine(Data); // perform halfpel refine of current best vector
 
-		if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(Data);
+			if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
+				Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
+				Data->currentQMV->y = 2 * Data->currentMV->y;
+			}
+		}
+
+		if(pParam->m_quarterpel) {
+			if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) &&
+				(MotionFlags & PMV_QUARTERPELREFINE8)) {
+			CheckCandidate = CheckCandidate8_qpel;
+			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
+				pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel);
+			QuarterpelRefine(Data);
+			}
+		}
+	}
+
+	if(pParam->m_quarterpel) {
+		pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x;
+		pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y;
+	}
+	else {
+		pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
+		pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
 	}
 
-	pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
-	pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
 	pMB->mvs[block] = *(Data->currentMV);
-	pMB->sad8[block] =  4 * (*Data->iMinSAD);
+	pMB->qmvs[block] = *(Data->currentQMV);
 
+	pMB->sad8[block] =  4 * (*Data->iMinSAD);
 }
 
 /* B-frames code starts here */
@@ -972,7 +1316,7 @@
 	Data->predMV = *predMV;
 
 	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
-				pParam->width, pParam->height, iFcode);
+				pParam->width, pParam->height, iFcode, pParam->m_quarterpel);
 
 	pmv[0] = Data->predMV;
 	PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
@@ -999,17 +1343,16 @@
 
 // three bits are needed to code backward mode. four for forward
 // we treat the bits just like they were vector's
-	if (mode_current == MODE_FORWARD) *Data->iMinSAD +=  4 * lambda_vec16[Data->iQuant];
-	else *Data->iMinSAD +=  3 * lambda_vec16[Data->iQuant];
-
+	if (mode_current == MODE_FORWARD) *Data->iMinSAD +=  4 * Data->lambda16;
+	else *Data->iMinSAD +=  3 * Data->lambda16;
 
 	if (*Data->iMinSAD < *best_sad) {
 		*best_sad = *Data->iMinSAD;
 		pMB->mode = mode_current;
 		pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
 		pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
-		if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
-		else pMB->b_mvs[0] = *Data->currentMV;
+		if (mode_current == MODE_FORWARD) pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV;
+		else pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
 	}
 	
 }
@@ -1086,8 +1429,8 @@
 	(*CheckCandidate)(0, 0, 255, &k, Data);
 
 // skip decision
-	if (*Data->iMinSAD - 2 * lambda_vec16[Data->iQuant] < (int32_t)Data->iQuant * SKIP_THRESH_B) {
-		//checking chroma. everything copied from MC
+	if (*Data->iMinSAD < pMB->quant * SKIP_THRESH_B) {
+		//possible skip - checking chroma. everything copied from MC
 		//this is not full chroma compensation, only it's fullpel approximation. should work though
 		int sum, dx, dy, b_dx, b_dy;
 
@@ -1112,7 +1455,7 @@
 					b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2,
 					Data->iEdgedWidth/2);
 
-		if ((uint32_t) sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) {
+		if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
 			pMB->mode = MODE_DIRECT_NONE_MV;
 			return *Data->iMinSAD; 
 		}
@@ -1131,7 +1474,7 @@
 
 	HalfpelRefine(Data);
 
-	*Data->iMinSAD +=  1 * lambda_vec16[Data->iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's
+	*Data->iMinSAD +=  1 * Data->lambda16; // one bit is needed to code direct mode
 	*best_sad = *Data->iMinSAD;
 
 	if (b_mb->mode == MODE_INTER4V) 
@@ -1187,12 +1530,11 @@
 	int iDirection, i, j;
 	SearchData bData;
 
-	bData.iMinSAD = fData->iMinSAD;
-	*bData.iMinSAD = 4096*256;
+	*(bData.iMinSAD = fData->iMinSAD) = 4096*256;
 	bData.Cur = fData->Cur;
 	fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth;
 	bData.currentMV = fData->currentMV + 1;
-	bData.iQuant = fData->iQuant;
+	bData.lambda16 = fData->lambda16;
 	fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
 
 	bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16;
@@ -1207,10 +1549,9 @@
 	bData.bpredMV = fData->predMV = *f_predMV;
 	fData->bpredMV = bData.predMV = *b_predMV;
 
-	fData->currentMV[0] = pMB->mvs[0];
-	fData->currentMV[1] = pMB->b_mvs[0];
-	get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode);
-	get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode);
+	fData->currentMV[0] = fData->currentMV[3]; //forward search stored it's vector here. backward stored it in the place it's needed
+	get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, pParam->m_quarterpel);
+	get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel);
 
 	if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
 	if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy;
@@ -1247,8 +1588,8 @@
 
 	} while (!(iDirection));
 
-// two bits are needed to code interpolate mode. we treat the bits just like they were vector's
-	*fData->iMinSAD +=  2 * lambda_vec16[fData->iQuant];
+	*fData->iMinSAD +=  2 * fData->lambda16; // two bits are needed to code interpolate mode.
+
 	if (*fData->iMinSAD < *best_sad) {
 		*best_sad = *fData->iMinSAD;
 		pMB->mvs[0] = fData->currentMV[0];
@@ -1262,7 +1603,6 @@
 	}
 }
 
-
 void
 MotionEstimationBVOP(MBParam * const pParam,
 					 FRAMEINFO * const frame,
@@ -1299,7 +1639,7 @@
 	Data.iEdgedWidth = pParam->edged_width;
 	Data.currentMV = currentMV;
 	Data.iMinSAD = &iMinSAD;
-	Data.iQuant = frame->quant;
+	Data.lambda16 = lambda_vec16[frame->quant];
 
 	// note: i==horizontal, j==vertical
 
@@ -1311,16 +1651,17 @@
 			MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
 			const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
 
-/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */
+/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
 			if (b_mb->mode == MODE_NOT_CODED) {
 				pMB->mode = MODE_NOT_CODED;
 				continue;
 			}
 
 			Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
+			pMB->quant = frame->quant;
+
 /* direct search comes first, because it (1) checks for SKIP-mode
 	and (2) sets very good predictions for forward and backward search */
-
 			skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
 									b_ref, b_refH->y, b_refV->y, b_refHV->y,
 									&frame->image,
@@ -1334,9 +1675,6 @@
 
 			if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
 
-//			best_sad = 256*4096; //uncomment to disable Directsearch.
-//	To disable any other mode, just comment the function call
-
 			// forward search
 			SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y,
 						&frame->image, i, j,
@@ -1389,72 +1727,12 @@
 			}
 		}
 	}
-
-//	fprintf(debug,"B-Stat: F: %04d   B: %04d   I: %04d  D: %04d, N: %04d\n",
-//				f_count,b_count,i_count,d_count,n_count);
-
 }
 
 /* Hinted ME starts here */
 
-static __inline void 
-Search8hinted(	const SearchData * const OldData,
-				const int x, const int y,
-				const uint32_t MotionFlags,
-				const MBParam * const pParam,
-				MACROBLOCK * const pMB,
-				const MACROBLOCK * const pMBs,
-				const int block)
-{
-	SearchData Data;
-	MainSearchFunc *MainSearchPtr;
-
-	Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block);
-	Data.iMinSAD = OldData->iMinSAD + 1 + block;
-	Data.currentMV = OldData->currentMV+1+block;
-	Data.iFcode = OldData->iFcode;
-	Data.iQuant = OldData->iQuant;
-
-	Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1));
-	Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1));
-	Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1));
-	Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1));
-	Data.iEdgedWidth = pParam->edged_width;
-	Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1));
-
-	CheckCandidate = CheckCandidate8;
-
-	if (block != 0)
-		*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * 
-								d_mv_bits(	Data.currentMV->x - Data.predMV.x, 
-											Data.currentMV->y - Data.predMV.y, 
-											Data.iFcode);
-
-
-	get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8,
-				pParam->width, pParam->height, OldData->iFcode);
-
-	if (pMB->mode == MODE_INTER4V) {
-		int dummy;
-		CheckCandidate8(pMB->mvs[block].x, pMB->mvs[block].y, 0, &dummy, &Data); }
-
-	if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
-		else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
-			else MainSearchPtr = DiamondSearch;
-
-	(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255);
-
-	if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data);
-
-	pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x;
-	pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y;
-	pMB->mvs[block] = *(Data.currentMV);
-	pMB->sad8[block] =  4 * (*(Data.iMinSAD));
-}
-
-
 static void
-SearchPhinted (	const uint8_t * const pRef,
+SearchPhinted (	const IMAGE * const pRef,
 				const uint8_t * const pRefH,
 				const uint8_t * const pRefV,
 				const uint8_t * const pRefHV,
@@ -1470,21 +1748,24 @@
 				SearchData * const Data)
 {
 
-	const int32_t iEdgedWidth = pParam->edged_width;
- 
 	int i, t;
 	MainSearchFunc * MainSearchPtr;
 
 	Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0);
+	Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
 	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
-				pParam->width, pParam->height, Data->iFcode);
+				pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel);
 
-	Data->Cur = pCur->y + (x + y * iEdgedWidth) * 16;
-	Data->Ref = pRef + (x + iEdgedWidth*y)*16;
-	Data->RefH = pRefH + (x + iEdgedWidth*y) * 16;
-	Data->RefV = pRefV + (x + iEdgedWidth*y) * 16;
-	Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16;
-	Data->iQuant = iQuant;
+	Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16;
+	Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8;
+	Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8;
+
+	Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16;
+	Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16;
+	Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16;
+	Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16;
+	Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
+	Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
 
 	if (!(MotionFlags & PMV_HALFPEL16)) {
 		Data->min_dx = EVEN(Data->min_dx);
@@ -1497,11 +1778,9 @@
 
 	if (pMB->dquant != NO_CHANGE) inter4v = 0;
 
-	if (inter4v)
-		CheckCandidate = CheckCandidate16;
+	if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16;
 	else CheckCandidate = CheckCandidate16no4v;
 
-
 	pMB->mvs[0].x = EVEN(pMB->mvs[0].x);
 	pMB->mvs[0].y = EVEN(pMB->mvs[0].y);
 	if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed
@@ -1529,26 +1808,71 @@
 
 	if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data);
 
-	if (inter4v)
-		for(i = 0; i < 4; i++)
-			Search8hinted(Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i);
+	for(i = 0; i < 5; i++) {
+		Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
+		Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
+	}
+
+	if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) {
+		get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
+				pParam->width, pParam->height, Data->iFcode, 0);
+		CheckCandidate = CheckCandidate16_qpel;
+		QuarterpelRefine(Data);
+	}
+
+	if (inter4v) {
+		SearchData Data8;
+		Data8.iFcode = Data->iFcode;
+		Data8.lambda8 = Data->lambda8;
+		Data8.iEdgedWidth = Data->iEdgedWidth;
+		Data8.RefQ = Data->RefQ;
+		Data8.qpel = Data->qpel;
+		Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
+		Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
+		Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
+		Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
+
+		if (Data->chroma) {
+			int sum, dx, dy;
+
+			if(pParam->m_quarterpel)
+				sum = (pMB->qmvs[0].y + pMB->qmvs[1].y + pMB->qmvs[2].y + pMB->qmvs[3].y)/2;
+			else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y;
+			dy = (sum ? SIGN(sum) *
+				  (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0);
+
+			if(pParam->m_quarterpel)
+				sum = (pMB->qmvs[0].x + pMB->qmvs[1].x + pMB->qmvs[2].x + pMB->qmvs[3].x)/2;
+			else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x;
+			dx = (sum ? SIGN(sum) *
+				  (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0);
+			Data->iMinSAD[1] += ChromaSAD(dx, dy, Data);
+		}
+	}
 
 	if (!(inter4v) ||
 		(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + 
 							Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) {
 // INTER MODE
-
 		pMB->mode = MODE_INTER;
 		pMB->mvs[0] = pMB->mvs[1]
 			= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
 
+		pMB->qmvs[0] = pMB->qmvs[1]
+			= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
+
 		pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] =
 			pMB->sad8[2] = pMB->sad8[3] =  Data->iMinSAD[0];
 
-		pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
-		pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
+		if(pParam->m_quarterpel) {
+			pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x;
+			pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y;
+		} else {
+			pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
+			pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
+		}
 	} else {
-// INTER4V MODE; all other things are already set in Search8hinted
+// INTER4V MODE; all other things are already set in Search8
 		pMB->mode = MODE_INTER4V;
 		pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] 
 						+ Data->iMinSAD[4] + IMV16X16 * iQuant;
@@ -1569,15 +1893,27 @@
 	const IMAGE *const pRef = &reference->image;
 
 	uint32_t x, y;
-	int32_t temp[5];
+	uint8_t * qimage;
+	int32_t temp[5], quant = current->quant;
 	int32_t iMinSAD[5];
-	VECTOR currentMV[5];
+	VECTOR currentMV[5], currentQMV[5];
 	SearchData Data;
 	Data.iEdgedWidth = pParam->edged_width;
 	Data.currentMV = currentMV;
+	Data.currentQMV = currentQMV;
 	Data.iMinSAD = iMinSAD;
 	Data.temp = temp;
 	Data.iFcode = current->fcode;
+	Data.rounding = pParam->m_rounding_type;
+	Data.qpel = pParam->m_quarterpel;
+	Data.chroma = current->global_flags & XVID_ME_COLOUR;
+
+	if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL)
+		return; // allocate some mem for qpel interpolated blocks
+				  // somehow this is dirty since I think we shouldn't use malloc outside
+				  // encoder_create() - so please fix me!
+
+	Data.RefQ = qimage;
 
 	if (sadInit) (*sadInit) ();
 
@@ -1592,14 +1928,23 @@
 			if (!(current->global_flags & XVID_LUMIMASKING)) {
 				pMB->dquant = NO_CHANGE;
 				pMB->quant = current->quant; }
+			else {
+				if (pMB->dquant != NO_CHANGE) {
+					quant += DQtab[pMB->dquant];
+					if (quant > 31) quant = 31;
+					else if (quant < 1) quant = 1;
+				}
+				pMB->quant = quant;
+			}
 
-			SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
+			SearchPhinted(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
 							y, current->motion_flags, pMB->quant,
 							pParam, pMBs, current->global_flags & XVID_INTER4V, pMB,
 							&Data);
 
 		}
 	}
+	free(qimage);
 }
 
 static __inline int
@@ -1613,13 +1958,13 @@
 				SearchData * const Data)
 {
 
-	int i, mask;
+	int i = 255, mask;
 	VECTOR pmv[3];
 
 	*(Data->iMinSAD) = MV_MAX_ERROR;
 	Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0);
 	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
-				pParam->width, pParam->height, Data->iFcode);
+				pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel);
 
 	Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
 	Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
@@ -1640,8 +1985,8 @@
 
 	DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
 
-	pMB->mvs[0] = pMB->mvs[1]
-			= pMB->mvs[2] = pMB->mvs[3] = *Data->currentMV; // all, for future get_pmv()
+	pMB->mvs[0] = *Data->currentMV;
+	pMB->mode = MODE_INTER;
 
 	return *(Data->iMinSAD);
 }
@@ -1666,28 +2011,24 @@
 	Data.currentMV = &currentMV;
 	Data.iMinSAD = &iMinSAD;
 	Data.iFcode = iFcode;
-	Data.iQuant = 2;
 
 	if (sadInit) (*sadInit) ();
 
-	for (y = 0; y < pParam->mb_height-1; y++) {
-		for (x = 0; x < pParam->mb_width; x++) {
+	for (y = 1; y < pParam->mb_height-1; y++) {
+		for (x = 1; x < pParam->mb_width-1; x++) {
 			int sad, dev;
 			MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
 
 			sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y,
 								pParam, pMBs, pMB, &Data);
 			
-			if ( x != 0 && y != 0 && x != pParam->mb_width-1 ) { //no edge macroblocks, they just don't work
-				if (sad > INTRA_THRESH) {
-					dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16,
-								  pParam->edged_width);
-					if (dev + INTRA_THRESH < sad) intra++; 
-					if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2;  // I frame
+			if (sad > INTRA_THRESH) {
+				dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16,
+							  pParam->edged_width);
+				if (dev + INTRA_THRESH < sad) { intra++; pMB->mode = MODE_INTRA; }
+				if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2;  // I frame
 				}
-				sSAD += sad;
-			}
-
+			sSAD += sad;
 		}
 	}
 	sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2);
@@ -1696,3 +2037,125 @@
 	return 0; // B frame
 
 }
+
+int
+FindFcode(	const MBParam * const pParam,
+			const FRAMEINFO * const current)
+{
+	uint32_t x, y;
+	int max = 0, min = 0, i;
+
+	for (y = 0; y < pParam->mb_height; y++) {
+		for (x = 0; x < pParam->mb_width; x++) {
+		
+			MACROBLOCK *pMB = &current->mbs[x + y * pParam->mb_width];
+			for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) {
+				if (pMB->mvs[i].x > max) max = pMB->mvs[i].x;
+				if (pMB->mvs[i].y > max) max = pMB->mvs[i].y;
+
+				if (pMB->mvs[i].x < min) min = pMB->mvs[i].x;
+				if (pMB->mvs[i].y < min) min = pMB->mvs[i].y;
+			}
+		}
+	}
+
+	min = -min;
+	max += 1;
+	if (min > max) max = min;
+	if (pParam->m_quarterpel) max *= 2;
+
+	for (i = 1; (max > 32 << (i - 1)); i++);
+	return i;
+}
+
+static void
+CheckGMC(int x, int y, const int dir, int * iDirection, 
+		const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC,
+		const MBParam * const pParam)
+{
+	uint32_t mx, my, a, count = 0;
+
+	for (my = 1; my < pParam->mb_height-1; my++)
+		for (mx = 1; mx < pParam->mb_width-1; mx++) {
+			VECTOR mv;
+			const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width];
+			if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue;
+			mv = pMB->mvs[0];
+			a = ABS(mv.x - x) + ABS(mv.y - y);
+			if (a < 6) count += 6 - a;
+		}
+
+	if (count > *bestcount) {
+		*bestcount = count;
+		*iDirection = dir;
+		GMC->x = x; GMC->y = y; 
+	}
+}
+
+
+static VECTOR
+GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode)
+{
+
+	uint32_t count, bestcount = 0;
+	int x, y;
+	VECTOR gmc = {0,0};
+	int step, min_x, max_x, min_y, max_y;
+	uint32_t mx, my;
+	int iDirection, bDirection;
+
+	min_x = min_y = -32<<iFcode;
+	max_x = max_y = 32<<iFcode;
+
+//step1: let's find a rough camera panning
+	for (step = 32; step >= 2; step /= 2) {
+		bestcount = 0;
+		for (y = min_y; y <= max_y; y += step)
+			for (x = min_x ; x <= max_x; x += step) {
+				count = 0;
+				//for all macroblocks
+				for (my = 1; my < pParam->mb_height-1; my++)
+					for (mx = 1; mx < pParam->mb_width-1; mx++) {
+						const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width];
+						VECTOR mv;
+						
+						if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) 
+							continue;
+	
+						mv = pMB->mvs[0];
+						if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) 	/* GMC translation is always halfpel-res */
+							count++;
+					}
+				if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; }
+			}
+		min_x = gmc.x - step;
+		max_x = gmc.x + step;
+		min_y = gmc.y - step;
+		max_y = gmc.y + step;
+
+	}
+	
+	if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) 
+		gmc.x = gmc.y = 0; //no camara pan, no GMC
+
+// step2: let's refine camera panning using gradiend-descent approach. 
+// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond)
+	bestcount = 0;
+	CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam);
+	do {
+		x = gmc.x; y = gmc.y;
+		bDirection = iDirection; iDirection = 0;
+		if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
+		if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
+		if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam);
+		if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
+
+	} while (iDirection);
+
+	if (pParam->m_quarterpel) {
+		gmc.x *= 2;
+		gmc.y *= 2;	/* we store the halfpel value as pseudo-qpel to make comparison easier */
+	}
+
+	return gmc;
+}