--- branches/dev-api-4/xvidcore/src/motion/motion_est.c	2003/08/28 11:14:04	1133
+++ branches/dev-api-4/xvidcore/src/motion/motion_est.c	2003/09/06 11:24:50	1138
@@ -21,7 +21,7 @@
  *  along with this program ; if not, write to the Free Software
  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  *
- * $Id: motion_est.c,v 1.58.2.30 2003-08-28 11:14:04 syskin Exp $
+ * $Id: motion_est.c,v 1.58.2.35 2003-09-06 11:24:50 Isibaar Exp $
  *
  ****************************************************************************/
 
@@ -347,6 +347,65 @@
 }
 
 static void
+CheckCandidate16_subpel(const int x, const int y, const SearchData * const data, const int Direction)
+{
+	int xc, yc;
+	const uint8_t *Reference;
+	VECTOR *current, *current2;
+	int32_t sad; uint32_t t;
+
+	if ( (x > data->max_dx) || (x < data->min_dx)
+		|| (y > data->max_dy) || (y < data->min_dy) ) return;
+
+	if (!data->qpel_precision) {
+		Reference = GetReference(x, y, data);
+		current = data->currentMV;
+		current2 = data->currentMV2;
+		xc = x; yc = y;
+	} else { /* x and y are in 1/4 precision */
+		Reference = Interpolate16x16qpel(x, y, 0, data);
+		xc = x/2; yc = y/2; /* for chroma sad */
+		current = data->currentQMV;
+		current2 = data->currentQMV2;
+	}
+
+	sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp);
+	t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
+
+	sad += (data->lambda16 * t * sad)>>10;
+	data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10;
+
+	if (data->chroma && sad < data->iMinSAD[0])
+		sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
+							(yc >> 1) + roundtab_79[yc & 0x3], data);
+
+	if (data->temp[0] < data->iMinSAD[1]) {
+		data->iMinSAD[1] = data->temp[0]; current[1].x = x; current[1].y = y; }
+	if (data->temp[1] < data->iMinSAD[2]) {
+		data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
+	if (data->temp[2] < data->iMinSAD[3]) {
+		data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
+	if (data->temp[3] < data->iMinSAD[4]) {
+		data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
+
+	if (sad < data->iMinSAD[0]) {
+		*(data->iMinSAD2) = *(data->iMinSAD);
+		current2->x = current->x; current2->y = current->y;
+
+		data->iMinSAD[0] = sad;
+		current[0].x = x; current[0].y = y;
+		*data->dir = Direction;
+		return;
+	}
+
+	if (sad < *(data->iMinSAD2)) {
+		*(data->iMinSAD2) = sad;
+		current2->x = x; current2->y = y;
+		*data->dir = Direction;
+	}
+}
+
+static void
 CheckCandidate8(const int x, const int y, const SearchData * const data, const int Direction)
 {
 	int32_t sad; uint32_t t;
@@ -487,7 +546,7 @@
 	if ( (x > data->max_dx) || (x < data->min_dx)
 		|| (y > data->max_dy) || (y < data->min_dy) ) return;
 
-	sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*((int)data->iEdgedWidth),
+	sad = sad32v_c(data->Cur, data->RefP[0] + x + y*((int)data->iEdgedWidth),
 					data->iEdgedWidth, data->temp);
 
 	if (sad < *(data->iMinSAD)) {
@@ -918,6 +977,106 @@
 /* MAINSEARCH FUNCTIONS END */
 
 static void
+SubpelRefine_Fast(SearchData * data, CheckFunc * CheckCandidate)
+{
+/* Do a half-pel or q-pel refinement */
+	VECTOR centerMV;
+	VECTOR second_best;
+	int best_sad = *data->iMinSAD;
+	int xo, yo, xo2, yo2;
+	int size = 2;
+	CheckFunc *backupFunc = CheckCandidate;
+
+	if(data->qpel_precision)
+		size = 1;
+
+	centerMV = *data->currentMV;
+	*data->iMinSAD = 256 * 4096;
+
+	CHECK_CANDIDATE(centerMV.x, centerMV.y - size, 0);
+	CHECK_CANDIDATE(centerMV.x + size, centerMV.y - size, 0);
+	CHECK_CANDIDATE(centerMV.x + size, centerMV.y, 0);
+	CHECK_CANDIDATE(centerMV.x + size, centerMV.y + size, 0);
+
+	CHECK_CANDIDATE(centerMV.x, centerMV.y + size, 0);
+	CHECK_CANDIDATE(centerMV.x - size, centerMV.y + size, 0);
+	CHECK_CANDIDATE(centerMV.x - size, centerMV.y, 0);
+	CHECK_CANDIDATE(centerMV.x - size, centerMV.y - size, 0);
+
+	second_best = *data->currentMV;
+	
+	if(data->qpel_precision) {
+		second_best.x *= 2;	second_best.y *= 2;
+	}
+
+	data->currentMV[0] = centerMV;
+	*data->iMinSAD = best_sad;
+
+    centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
+
+	xo = centerMV.x;
+	yo = centerMV.y;
+	xo2 = second_best.x;
+	yo2 = second_best.y;
+
+	CheckCandidate = CheckCandidate16_subpel;
+	*data->iMinSAD2 = 256 * 4096;
+
+	if (yo == yo2)
+	{
+		CHECK_CANDIDATE((xo+xo2)>>1, yo, 0);
+		CHECK_CANDIDATE(xo, yo-1, 0);
+		CHECK_CANDIDATE(xo, yo+1, 0);
+		
+		if(best_sad <= *data->iMinSAD2)
+			goto ende;
+
+		if(data->currentQMV[0].x == data->currentQMV2[0].x) {
+			CHECK_CANDIDATE((xo+xo2)>>1, yo-1, 0);
+			CHECK_CANDIDATE((xo+xo2)>>1, yo+1, 0);
+			goto ende;
+		}
+		else {
+			CHECK_CANDIDATE((xo+xo2)>>1,
+				(data->currentQMV[0].x == xo) ? data->currentQMV[0].y : data->currentQMV2[0].y,
+				0);
+			goto ende;
+		}
+	} 
+
+	if (xo == xo2)
+	{
+		CHECK_CANDIDATE(xo, (yo+yo2)>>1, 0);
+		CHECK_CANDIDATE(xo-1, yo, 0);
+		CHECK_CANDIDATE(xo+1, yo, 0);
+		
+		if(best_sad < *data->iMinSAD2)
+			goto ende;
+
+		if(data->currentQMV[0].y == data->currentQMV2[0].y) {
+			CHECK_CANDIDATE(xo-1, (yo+yo2)>>1, 0);
+			CHECK_CANDIDATE(xo+1, (yo+yo2)>>1, 0);
+			goto ende;
+		}
+		else {
+			CHECK_CANDIDATE((data->currentQMV[0].y == yo) ? data->currentQMV[0].x : data->currentQMV2[0].x, (yo+yo2)>>1, 0);
+			goto ende;
+		}
+	}
+
+	CHECK_CANDIDATE(xo, (yo+yo2)>>1, 0);
+	CHECK_CANDIDATE((xo+xo2)>>1, yo, 0);
+
+	if(best_sad <= *data->iMinSAD2)
+		goto ende;
+
+	CHECK_CANDIDATE((xo+xo2)>>1, (yo+yo2)>>1, 0);
+
+ende:
+	CheckCandidate = backupFunc;
+}
+
+static void
 SubpelRefine(const SearchData * const data, CheckFunc * const CheckCandidate)
 {
 /* Do a half-pel or q-pel refinement */
@@ -970,6 +1129,265 @@
 }
 
 static __inline void
+ModeDecision_Fast(SearchData * const Data,
+			MACROBLOCK * const pMB,
+			const MACROBLOCK * const pMBs,
+			const int x, const int y,
+			const MBParam * const pParam,
+			const uint32_t MotionFlags,
+			const uint32_t VopFlags,
+			const uint32_t VolFlags,
+			const IMAGE * const pCurrent,
+			const IMAGE * const pRef,
+			const IMAGE * const vGMC,
+			const int coding_type)
+{
+	int mode = MODE_INTER;
+	int mcsel = 0;
+	int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
+	const uint32_t iQuant = pMB->quant;
+	const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0);
+    int sad;
+	int min_rd = -1, intra_rd, i, cbp, c[2] = {0, 0};
+	VECTOR backup[5], *v;
+	int sad_backup[5];
+	int InterBias = MV16_INTER_BIAS;
+	int thresh = 0;
+	int count = 0;
+	int top = 0, top_right = 0, left = 0;
+
+	pMB->mcsel = 0;
+
+	/* INTER <-> INTER4V decision */
+	if ((Data->iMinSAD[0] + 75 < Data->iMinSAD[1] + 
+		Data->iMinSAD[2] + Data->iMinSAD[3] + Data->iMinSAD[4])) { /* normal, fast, SAD-based mode decision */
+		if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
+			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
+			mode = MODE_INTER;
+			sad = Data->iMinSAD[0];
+		} else {
+			mode = MODE_INTER4V;
+			sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
+						Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
+			Data->iMinSAD[0] = sad;
+		}
+
+		/* final skip decision, a.k.a. "the vector you found, really that good?" */
+		if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP))
+			if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH)
+				if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) {
+					mode = MODE_NOT_CODED;
+					goto early_out;
+				}
+
+		/* mcsel */
+		if (coding_type == S_VOP) {
+
+			int32_t iSAD = sad16(Data->Cur,
+				vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536);
+
+			if (Data->chroma) {
+				iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
+				iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
+			}
+
+			if (iSAD <= sad) {		/* mode decision GMC */
+				mode = MODE_INTER;
+				mcsel = 1;
+				sad = iSAD;
+			}
+
+		}
+	} else { /* Rate-Distortion INTER<->INTER4V */
+		Data->iQuant = iQuant;
+		Data->cbp = c;
+		v = Data->qpel ? Data->currentQMV : Data->currentMV;
+
+		/* final skip decision, a.k.a. "the vector you found, really that good?" */
+		if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP))
+			if ( (100*Data->iMinSAD[0])/(pMB->sad16+1) > FINAL_SKIP_THRESH)
+				if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) {
+					mode = MODE_NOT_CODED;
+					goto early_out;
+				}
+
+		for (i = 0; i < 5; i++) {
+			sad_backup[i] = Data->iMinSAD[i];
+			Data->iMinSAD[i] = 256*4096;
+			backup[i] = v[i];
+		}
+
+		min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags);
+		cbp = *Data->cbp;
+		sad = sad_backup[0];
+
+		if (coding_type == S_VOP) {
+			int gmc_rd;
+			int32_t iSAD = sad16(Data->Cur,
+				vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536);
+
+			if (Data->chroma) {
+				iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
+				iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
+			}
+
+			*Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */
+			gmc_rd = findRDgmc(Data, vGMC, x, y);
+			if (gmc_rd < min_rd) {
+				mcsel = 1;
+				*Data->iMinSAD = min_rd = gmc_rd;
+				mode = MODE_INTER;
+				cbp = *Data->cbp;
+				sad = iSAD;
+			}
+		}
+
+		if (inter4v) {
+			int v4_rd;
+			v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
+			if (v4_rd < min_rd) {
+				Data->iMinSAD[0] = min_rd = v4_rd;
+				mode = MODE_INTER4V;
+				cbp = *Data->cbp;
+				sad = sad_backup[1] + sad_backup[2] +
+					  sad_backup[3] + sad_backup[4] + IMV16X16 * (int32_t)iQuant;
+			}
+		}
+	}
+
+	left = top = top_right = -1;
+	thresh = 0;
+
+	if((x > 0) && (y > 0) && (x < (int32_t) pParam->mb_width)) {
+		left = (&pMBs[(x-1) + y * pParam->mb_width])->sad16; // left
+		top = (&pMBs[x + (y-1) * pParam->mb_width])->sad16; // top
+		top_right = (&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16; // top right
+
+		if(((&pMBs[(x-1) + y * pParam->mb_width])->mode != MODE_INTRA) &&
+		   ((&pMBs[x + (y-1) * pParam->mb_width])->mode != MODE_INTRA) &&
+		   ((&pMBs[(x+1) + (y-1) * pParam->mb_width])->mode != MODE_INTRA)) {
+			thresh = MAX(MAX(top, left), top_right);
+		}
+		else
+			thresh = MIN(MIN(top, left), top_right);
+	}
+
+	/* INTRA <-> INTER decision */
+	if (sad < thresh) { /* normal, fast, SAD-based mode decision */
+		/* intra decision */
+
+		if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */
+		if (y != 0)
+			if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
+		if (x != 0)
+			if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
+
+		if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */
+		if (Data->rrv) InterBias *= 4;
+
+		if (InterBias < sad) {
+			int32_t deviation;
+			if (!Data->rrv)
+				deviation = dev16(Data->Cur, Data->iEdgedWidth);
+			else
+				deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */
+							dev16(Data->Cur+16, Data->iEdgedWidth) +
+							dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) +
+							dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth);
+
+			if (deviation < (sad - InterBias)) mode = MODE_INTRA;
+		}
+
+		pMB->cbp = 63;
+	} else { /* Rate-Distortion INTRA<->INTER */
+		if(min_rd < 0) {
+			Data->iQuant = iQuant;
+			Data->cbp = c;
+			v = Data->qpel ? Data->currentQMV : Data->currentMV;
+
+			for (i = 0; i < 5; i++) {
+				Data->iMinSAD[i] = 256*4096;
+				backup[i] = v[i];
+			}
+
+			if(mode == MODE_INTER) {
+				min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags);
+				cbp = *Data->cbp;
+
+				if (coding_type == S_VOP) {
+					int gmc_rd;
+
+					*Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */
+					gmc_rd = findRDgmc(Data, vGMC, x, y);
+					if (gmc_rd < min_rd) {
+						mcsel = 1;
+						*Data->iMinSAD = min_rd = gmc_rd;
+						mode = MODE_INTER;
+						cbp = *Data->cbp;
+					}
+				}
+			}
+
+			if(mode == MODE_INTER4V) {
+				int v4_rd;
+				v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
+				if (v4_rd < min_rd) {
+					Data->iMinSAD[0] = min_rd = v4_rd;
+					mode = MODE_INTER4V;
+					cbp = *Data->cbp;
+				}
+			}
+		}
+
+		intra_rd = findRDintra(Data);
+		if (intra_rd < min_rd) {
+			*Data->iMinSAD = min_rd = intra_rd;
+			mode = MODE_INTRA;
+		}
+
+		pMB->cbp = cbp;
+	}
+
+early_out:
+	pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
+
+	if (Data->rrv) {
+			Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
+			Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
+	}
+
+	if (mode == MODE_INTER && mcsel == 0) {
+		pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
+
+		if(Data->qpel) {
+			pMB->qmvs[0] = pMB->qmvs[1]
+				= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
+			pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
+			pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.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 if (mode == MODE_INTER ) { // but mcsel == 1
+
+		pMB->mcsel = 1;
+		if (Data->qpel) {
+			pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv;
+			pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2;
+			pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2;
+		} else
+			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv;
+
+	} else
+		if (mode == MODE_INTER4V) ; /* anything here? */
+	else	/* INTRA, NOT_CODED */
+		ZeroMacroblockP(pMB, 0);
+
+	pMB->mode = mode;
+}
+
+static __inline void
 ModeDecision(SearchData * const Data,
 			MACROBLOCK * const pMB,
 			const MACROBLOCK * const pMBs,
@@ -1162,6 +1580,7 @@
 	uint32_t mb_height = pParam->mb_height;
 	const uint32_t iEdgedWidth = pParam->edged_width;
 	const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
+	int stat_thresh = 0;
 
 	uint32_t x, y;
 	uint32_t iIntra = 0;
@@ -1174,14 +1593,20 @@
 	int32_t temp[8]; uint32_t dir;
 	VECTOR currentMV[5];
 	VECTOR currentQMV[5];
+	VECTOR currentMV2[5];
+	VECTOR currentQMV2[5];
 	int32_t iMinSAD[5];
+	int32_t iMinSAD2[5];
 	DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE);
 	SearchData Data;
 	memset(&Data, 0, sizeof(SearchData));
 	Data.iEdgedWidth = iEdgedWidth;
 	Data.currentMV = currentMV;
 	Data.currentQMV = currentQMV;
+	Data.currentMV2 = currentMV2;
+	Data.currentQMV2 = currentQMV2;
 	Data.iMinSAD = iMinSAD;
+	Data.iMinSAD2 = iMinSAD2;
 	Data.temp = temp;
 	Data.dir = &dir;
 	Data.iFcode = current->fcode;
@@ -1204,6 +1629,7 @@
 	for (y = 0; y < mb_height; y++)	{
 		for (x = 0; x < mb_width; x++)	{
 			MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
+			MACROBLOCK *prevMB = &reference->mbs[x + y * pParam->mb_width];
 
 			if (!Data.rrv) pMB->sad16 =
 				sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
@@ -1236,8 +1662,26 @@
 					}
 			}
 
+			if(MotionFlags & XVID_ME_DETECT_STATIC_MOTION) {
+				if(x > 0 && y > 0 && x < pParam->mb_width)
+					if(MVequal((&pMBs[(x-1) + y * pParam->mb_width])->mvs[0], zeroMV) &&
+					   MVequal((&pMBs[x + (y-1) * pParam->mb_width])->mvs[0], zeroMV) &&
+				       MVequal((&pMBs[(x+1) + (y-1) * pParam->mb_width])->mvs[0], zeroMV) &&
+				       MVequal(prevMB->mvs[0], zeroMV)) {
+						stat_thresh = MAX((&pMBs[(x-1) + y * pParam->mb_width])->sad16,
+								 	  MAX((&pMBs[x + (y-1) * pParam->mb_width])->sad16,
+									  MAX((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16,
+									  prevMB->sad16)));
+					}
+				else
+					stat_thresh = MIN((&pMBs[(x-1) + y * pParam->mb_width])->sad16,
+								 	  MIN((&pMBs[x + (y-1) * pParam->mb_width])->sad16,
+									  MIN((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16,
+									  prevMB->sad16)));
+			}
+
 			if ((current->vop_flags & XVID_VOP_CARTOON) && 
-				(sad00 < pMB->quant * 4 * skip_thresh)) { /* favorize (0,0) vector for cartoons */
+				(sad00 < pMB->quant * 4 * skip_thresh) || (sad00 < stat_thresh)) { /* favorize (0,0) vector for cartoons */
 				ZeroMacroblockP(pMB, sad00);
 				continue;
 			}
@@ -1246,9 +1690,16 @@
 					y, MotionFlags, current->vop_flags, current->vol_flags,
 					&Data, pParam, pMBs, reference->mbs, pMB);
 
-			ModeDecision(&Data, pMB, pMBs, x, y, pParam,
-						 MotionFlags, current->vop_flags, current->vol_flags,
-						 pCurrent, pRef, pGMC, current->coding_type);
+			if(current->vop_flags & XVID_VOP_FAST_MODEDECISION_RD) {
+				ModeDecision_Fast(&Data, pMB, pMBs, x, y, pParam,
+							 MotionFlags, current->vop_flags, current->vol_flags,
+							 pCurrent, pRef, pGMC, current->coding_type);
+			}
+			else {
+				ModeDecision(&Data, pMB, pMBs, x, y, pParam,
+							 MotionFlags, current->vop_flags, current->vol_flags,
+							 pCurrent, pRef, pGMC, current->coding_type);
+			}
 
 			if (pMB->mode == MODE_INTRA)
 				if (++iIntra > iLimit) return 1;
@@ -1472,7 +1923,10 @@
 				pParam->width, pParam->height, Data->iFcode, 2, 0);
 		Data->qpel_precision = 1;
 		if (MotionFlags & XVID_ME_QUARTERPELREFINE16)
-			SubpelRefine(Data, CheckCandidate);
+			if(MotionFlags & XVID_ME_FASTREFINE16)
+				SubpelRefine_Fast(Data, CheckCandidate);
+			else
+				SubpelRefine(Data, CheckCandidate);
 	}
 
 	if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30)
@@ -1671,7 +2125,7 @@
 			SearchData * const Data)
 {
 
-	int i, mask;
+	int i;
 	VECTOR pmv[7];
 	MainSearchFunc *MainSearchPtr;
 	*Data->iMinSAD = MV_MAX_ERROR;
@@ -1707,9 +2161,10 @@
 	else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
 		else MainSearchPtr = DiamondSearch;
 
-	mask = make_mask(pmv, 7, *Data->dir);
-
-	MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v);
+	if (*Data->iMinSAD > 512) {
+		unsigned int mask = make_mask(pmv, 7, *Data->dir);
+		MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v);
+	}
 
 	SubpelRefine(Data, CheckCandidate16no4v);
 
@@ -1759,16 +2214,15 @@
 {
 	int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
 	int32_t sum;
-	const int div = 1 + Data->qpel;
 	int k;
 	const uint32_t stride = Data->iEdgedWidth/2;
 	/* this is not full chroma compensation, only it's fullpel approximation. should work though */
 
 	for (k = 0; k < 4; k++) {
-		dy += Data->directmvF[k].y / div;
-		dx += Data->directmvF[k].x / div;
-		b_dy += Data->directmvB[k].y / div;
-		b_dx += Data->directmvB[k].x / div;
+		dy += Data->directmvF[k].y >> Data->qpel;
+		dx += Data->directmvF[k].x >> Data->qpel;
+		b_dy += Data->directmvB[k].y >> Data->qpel;
+		b_dx += Data->directmvB[k].x >> Data->qpel;
 	}
 
 	dy = (dy >> 3) + roundtab_76[dy & 0xf];
@@ -1781,14 +2235,14 @@
 					b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
 					stride);
 
-	if (sum >= MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; /* no skip */
+	if (sum >= MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) return; /* no skip */
 
 	sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
 					f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
 					b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
 					stride);
 
-	if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
+	if (sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) {
 		pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */
 		for (k = 0; k < 4; k++) {
 			pMB->qmvs[k] = pMB->mvs[k];
@@ -1873,7 +2327,7 @@
 	CheckCandidate(0, 0, Data, 255);
 
 	/* initial (fast) skip decision */
-	if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) {
+	if (*Data->iMinSAD < Data->iQuant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) {
 		/* possible skip */
 		if (Data->chroma) {
 			pMB->mode = MODE_DIRECT_NONE_MV;
@@ -2102,12 +2556,13 @@
 	Data.iEdgedWidth = pParam->edged_width;
 	Data.currentMV = currentMV; Data.currentQMV = currentQMV;
 	Data.iMinSAD = &iMinSAD;
-	Data.lambda16 = lambda_vec16[frame->quant];
+	Data.lambda16 = lambda_vec16[MAX(frame->quant-2, 2)];
 	Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0;
 	Data.rounding = 0;
 	Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP;
 	Data.temp = temp;
 	Data.dir = &dir;
+	Data.iQuant = frame->quant;
 
 	Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */
 
@@ -2130,7 +2585,6 @@
 			Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
 			Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
 			Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
-			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 */
@@ -2176,7 +2630,7 @@
 						&Data);
 
 			/* final skip decision */
-			if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
+			if ( (skip_sad < Data.iQuant * MAX_SAD00_FOR_SKIP * 2)
 					&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
 				SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
 
@@ -2218,54 +2672,63 @@
 	VECTOR pmv[3];
 	MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
 
-	for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
+	unsigned int simplicity = 0;
 
-	/* median is only used as prediction. it doesn't have to be real */
-	if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
-	else
-		if (x == 1) /* left macroblock does not have any vector now */
-			Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */
-		else if (y == 1) /* top macroblock doesn't have it's vector */
-			Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */
-			else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */
+	for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
 
 	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
-			pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, 0);
+			pParam->width, pParam->height, Data->iFcode - Data->qpel - 1, 0, 0);
 
 	Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
 	Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16;
 
-	pmv[1].x = EVEN(pMB->mvs[0].x);
-	pmv[1].y = EVEN(pMB->mvs[0].y);
-	pmv[2].x = EVEN(Data->predMV.x);
-	pmv[2].y = EVEN(Data->predMV.y);
-	pmv[0].x = pmv[0].y = 0;
+	pmv[0].x = pMB->mvs[0].x;
+	pmv[0].y = pMB->mvs[0].y;
+
+	CheckCandidate32I(pmv[0].x, pmv[0].y, Data, 0);
+
+	if (*Data->iMinSAD > 200) {
 
-	CheckCandidate32I(0, 0, Data, 0);
+		pmv[1].x = pmv[1].y = 0;
 
-	if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) {
+		/* median is only used as prediction. it doesn't have to be real */
+		if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
+		else
+			if (x == 1) /* left macroblock does not have any vector now */
+				Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */
+			else if (y == 1) /* top macroblock doesn't have it's vector */
+				Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */
+			else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */
+
+		pmv[2].x = Data->predMV.x;
+		pmv[2].y = Data->predMV.y;
 
 		if (!vector_repeats(pmv, 1))
 			CheckCandidate32I(pmv[1].x, pmv[1].y, Data, 1);
 		if (!vector_repeats(pmv, 2))
 			CheckCandidate32I(pmv[2].x, pmv[2].y, Data, 2);
 
-		if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) { /* diamond only if needed */
+		if (*Data->iMinSAD > 500) { /* diamond only if needed */
 			unsigned int mask = make_mask(pmv, 3, *Data->dir);
-			DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, 255/*mask*/, CheckCandidate32I);
-		}
-	}
+			DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate32I);
+		} else simplicity++;
+
+		if (*Data->iMinSAD > 500) /* refinement from 2-pixel to 1-pixel */
+			SubpelRefine(Data, CheckCandidate32I);
+		else simplicity++;
+	} else simplicity++;
 
 	for (i = 0; i < 4; i++) {
 		MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
 		MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
 		MB->mode = MODE_INTER;
-		MB->sad16 = Data->iMinSAD[i+1];
+		/* if we skipped some search steps, we have to assume that SAD would be lower with them */
+		MB->sad16 = Data->iMinSAD[i+1] - (simplicity<<7);
 	}
 }
 
 #define INTRA_THRESH	2200
-#define INTER_THRESH	50
+#define INTER_THRESH	40
 #define INTRA_THRESH2	95
 
 int
@@ -2296,6 +2759,7 @@
 	Data.temp = temp;
 	Data.dir = &dir;
 	Data.qpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0;
+	Data.qpel_precision = 0;
 
 	if (intraCount != 0) {
 		if (intraCount < 10) // we're right after an I frame
@@ -2305,8 +2769,8 @@
 				IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra;
 	}
 
-	InterThresh -= 12 * bCount;
-	if (InterThresh < 15 + b_thresh) InterThresh = 15 + b_thresh;
+	InterThresh -= 20 * bCount;
+	if (InterThresh < 10 + b_thresh) InterThresh = 10 + b_thresh;
 
 	if (sadInit) (*sadInit) ();
 
@@ -2338,10 +2802,10 @@
 				}
 
 				if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0)
-					if (dev > 500 && pMB->sad16 < 1000)
+					if (dev > 1000 && pMB->sad16 < 1000)
 						sSAD += 1000;
 
-				sSAD += (dev < 3000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */
+				sSAD += (dev < 4000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */
 			}
 		}
 	}