--- trunk/xvidcore/src/motion/estimation_rd_based.c	2004/12/05 04:53:01	1564
+++ trunk/xvidcore/src/motion/estimation_rd_based.c	2004/12/18 12:06:43	1579
@@ -20,7 +20,7 @@
  *  along with this program ; if not, write to the Free Software
  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  *
- * $Id: estimation_rd_based.c,v 1.9 2004-12-05 04:53:01 syskin Exp $
+ * $Id: estimation_rd_based.c,v 1.12 2004-12-18 12:06:43 syskin Exp $
  *
  ****************************************************************************/
 
@@ -57,7 +57,8 @@
 				const int block,
 				const uint16_t * scan_table,
 				const unsigned int lambda,
-				const uint16_t * mpeg_quant_matrices)
+				const uint16_t * mpeg_quant_matrices,
+				const unsigned int quant_sq)
 {
 	int sum;
 	int bits;
@@ -93,7 +94,7 @@
 	}
 
 
-	return bits + (lambda*distortion)/(quant*quant);
+	return bits + (lambda*distortion)/quant_sq;
 }
 
 static __inline unsigned int
@@ -111,7 +112,8 @@
 					unsigned int bits[2],
 					unsigned int cbp[2],
 					unsigned int lambda,
-					const uint16_t * mpeg_quant_matrices)
+					const uint16_t * mpeg_quant_matrices,
+					const unsigned int quant_sq)
 {
 	int direction;
 	int16_t *pCurrent;
@@ -146,8 +148,8 @@
 	/* dc prediction */
 	qcoeff[0] = qcoeff[0] - predictors[0];
 
-	if (block < 4) bits[1] = bits[0] = dcy_tab[qcoeff[0] + 255].len;
-	else bits[1] = bits[0] = dcc_tab[qcoeff[0] + 255].len;
+	if (block < 4) bits[1] = bits[0] = dcy_tab[qcoeff[0] + 255].len - 3; /* 3 bits added before (4 times) */
+	else bits[1] = bits[0] = dcc_tab[qcoeff[0] + 255].len - 2; /* 2 bits added before (2 times)*/
 
 	/* calc cost before ac prediction */
 	bits[0] += coded = CodeCoeffIntra_CalcBits(qcoeff, scan_tables[0]);
@@ -171,7 +173,7 @@
 
 	distortion = sse8_16bit(coeff, dqcoeff, 8*sizeof(int16_t));
 
-	return (lambda*distortion)/(quant*quant);
+	return (lambda*distortion)/quant_sq;
 }
 
 
@@ -181,7 +183,8 @@
 {
 
 	int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
-	int32_t rd = 0;
+	/* minimum nuber of bits INTER can take is 1 (mcbpc) + 2 (cby) + 2 (vector) */
+	int32_t rd = BITS_MULT * (1+2+2);
 	VECTOR * current;
 	const uint8_t * ptr;
 	int i, t, xc, yc;
@@ -203,10 +206,12 @@
 	for(i = 0; i < 4; i++) {
 		int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
 		transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
-		rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i, data->scan_table, data->lambda[i], data->mpeg_quant_matrices);
+		rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, 
+								data->quant_type, &cbp, i, data->scan_table, data->lambda[i], 
+								data->mpeg_quant_matrices, data->quant_sq);
 	}
 
-	rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision);
+	rd += t = BITS_MULT * (d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision) - 2);
 
 	if (data->temp[0] + t < data->iMinSAD[1]) {
 		data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) | (cbp&32); }
@@ -217,7 +222,7 @@
 	if (data->temp[3] < data->iMinSAD[4]) {
 		data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | (cbp&4); }
 
-	rd += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
+	rd += BITS_MULT * (xvid_cbpy_tab[15-(cbp>>2)].len - 2);
 
 	if (rd >= data->iMinSAD[0]) return;
 
@@ -228,15 +233,19 @@
 	/* chroma U */
 	ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
 	transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2);
-	rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4, data->scan_table, data->lambda[4], data->mpeg_quant_matrices);
+	rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type,
+							&cbp, 4, data->scan_table, data->lambda[4], 
+							data->mpeg_quant_matrices, data->quant_sq);
 	if (rd >= data->iMinSAD[0]) return;
 
 	/* chroma V */
 	ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
 	transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2);
-	rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5, data->scan_table, data->lambda[5], data->mpeg_quant_matrices);
+	rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type,
+							&cbp, 5, data->scan_table, data->lambda[5], 
+							data->mpeg_quant_matrices, data->quant_sq);
 
-	rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
+	rd += BITS_MULT * (mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len - 1); /* one was added before */
 
 	if (rd < data->iMinSAD[0]) {
 		data->iMinSAD[0] = rd;
@@ -268,8 +277,12 @@
 	}
 
 	transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
-	rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5, data->scan_table, data->lambda[0], data->mpeg_quant_matrices);
-	rd += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision);
+	rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, 
+						&cbp, 5, data->scan_table, data->lambda[0], 
+						data->mpeg_quant_matrices, data->quant_sq);
+	/* we took 2 bits into account before */
+	rd += BITS_MULT * (d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision) - 2); 
+
 
 	if (rd < data->iMinSAD[0]) {
 		*data->cbp = cbp;
@@ -353,7 +366,10 @@
 				const VECTOR * const backup)
 {
 
-	unsigned int cbp = 0, bits = 0, t = 0, i;
+	unsigned int cbp = 0, t = 0, i;
+	
+	/* minimum number of bits INTER4V can take is 2 (cbpy) + 3 (mcbpc) + 4*2 (vectors)*/
+	int bits = (2+3+4*2)*BITS_MULT;
 	SearchData Data2, *Data8 = &Data2;
 	int sumx = 0, sumy = 0;
 	int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64;
@@ -377,17 +393,17 @@
 		if(Data->qpel) {
 			Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
 			if (i != 0)	t = d_mv_bits(	Data8->currentQMV->x, Data8->currentQMV->y,
-										Data8->predMV, Data8->iFcode, 0);
+										Data8->predMV, Data8->iFcode, 0) - 2;
 		} else {
 			Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
 			if (i != 0)	t = d_mv_bits(	Data8->currentMV->x, Data8->currentMV->y,
-										Data8->predMV, Data8->iFcode, 0);
+										Data8->predMV, Data8->iFcode, 0) - 2;
 		}
 
 		get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3,
 					pParam->width, pParam->height, Data8->iFcode, Data8->qpel+1);
 
-		*Data8->iMinSAD += BITS_MULT*t;
+		*Data8->iMinSAD += BITS_MULT * t;
 
 		Data8->qpel_precision = Data8->qpel;
 		/* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */
@@ -482,7 +498,7 @@
 
 	} /* end - for all luma blocks */
 
-	bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
+	bits += BITS_MULT * (xvid_cbpy_tab[15-(cbp>>2)].len - 2); /* 2 were added before */
 
 	/* let's check chroma */
 	sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
@@ -491,16 +507,18 @@
 	/* chroma U */
 	ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
 	transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
-	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, Data->scan_table, Data->lambda[4], Data->mpeg_quant_matrices);
+	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4,
+							Data->scan_table, Data->lambda[4], Data->mpeg_quant_matrices, Data->quant_sq);
 
 	if (bits >= *Data->iMinSAD) return bits;
 
 	/* chroma V */
 	ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
 	transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
-	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, Data->scan_table, Data->lambda[5], Data->mpeg_quant_matrices);
+	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5,
+							Data->scan_table, Data->lambda[5], Data->mpeg_quant_matrices, Data->quant_sq);
 
-	bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
+	bits += BITS_MULT*(mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len - 3); /* 3 were added before */
 
 	*Data->cbp = cbp;
 	return bits;
@@ -511,7 +529,8 @@
 			 const int x, const int y, const int mb_width)
 {
 	unsigned int cbp[2] = {0, 0}, bits[2], i;
-	unsigned int bits1 = BITS_MULT*1, bits2 = BITS_MULT*1; /* this one is ac/dc prediction flag bit */
+	/* minimum number of bits that WILL be coded in intra - mcbpc 5, cby 2 acdc flag - 1 and DC coeffs - 4*3+2*2 */
+	int bits1 = BITS_MULT*(5+2+1+4*3+2*2), bits2 = BITS_MULT*(5+2+1+4*3+2*2);
 	unsigned int distortion = 0;
 
 	int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64, * dqcoeff = Data->dctSpace + 128;
@@ -524,7 +543,8 @@
 		
 
 		distortion = Block_CalcBitsIntra(pMB, x, y, mb_width, i, in, coeff, dqcoeff,
-								predictors[i], iQuant, Data->quant_type, bits, cbp, Data->lambda[i], Data->mpeg_quant_matrices);
+								predictors[i], iQuant, Data->quant_type, bits, cbp, 
+								Data->lambda[i], Data->mpeg_quant_matrices, Data->quant_sq);
 		bits1 += distortion + BITS_MULT * bits[0];
 		bits2 += distortion + BITS_MULT * bits[1];
 
@@ -532,13 +552,14 @@
 			return bits1;
 	}
 
-	bits1 += BITS_MULT*xvid_cbpy_tab[cbp[0]>>2].len;
-	bits2 += BITS_MULT*xvid_cbpy_tab[cbp[1]>>2].len;
+	bits1 += BITS_MULT * (xvid_cbpy_tab[cbp[0]>>2].len - 2); /* two bits were added before */
+	bits2 += BITS_MULT * (xvid_cbpy_tab[cbp[1]>>2].len - 2);
 
 	/*chroma U */
 	transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
 	distortion = Block_CalcBitsIntra(pMB, x, y, mb_width, 4, in, coeff, dqcoeff,
-									predictors[4], iQuant, Data->quant_type, bits, cbp, Data->lambda[4], Data->mpeg_quant_matrices);
+									predictors[4], iQuant, Data->quant_type, bits, cbp, 
+									Data->lambda[4], Data->mpeg_quant_matrices, Data->quant_sq);
 	bits1 += distortion + BITS_MULT * bits[0];
 	bits2 += distortion + BITS_MULT * bits[1];
 
@@ -548,13 +569,14 @@
 	/* chroma V */
 	transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
 	distortion = Block_CalcBitsIntra(pMB, x, y, mb_width, 5, in, coeff, dqcoeff,
-									predictors[5], iQuant, Data->quant_type, bits, cbp, Data->lambda[5], Data->mpeg_quant_matrices);
+									predictors[5], iQuant, Data->quant_type, bits, cbp, 
+									Data->lambda[5], Data->mpeg_quant_matrices, Data->quant_sq);
 
 	bits1 += distortion + BITS_MULT * bits[0];
 	bits2 += distortion + BITS_MULT * bits[1];
 
-	bits1 += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp[0] & 3) << 3)].len;
-	bits2 += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp[1] & 3) << 3)].len;
+	bits1 += BITS_MULT * (mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp[0] & 3) << 3)].len - 5); /* 5 bits were added before */
+	bits2 += BITS_MULT * (mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp[1] & 3) << 3)].len - 5);
 
 	*Data->cbp = bits1 <= bits2 ? cbp[0] : cbp[1];
 
@@ -565,30 +587,34 @@
 static int
 findRD_gmc(SearchData * const Data, const IMAGE * const vGMC, const int x, const int y)
 {
-	int bits = BITS_MULT*1; /* this one is mcsel */
+	/* minimum nubler of bits - 1 (mcbpc) + 2 (cby) + 1 (mcsel) */
+	int bits = BITS_MULT * (1+2+1);
 	unsigned int cbp = 0, i;
 	int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
 
 	for(i = 0; i < 4; i++) {
 		int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
 		transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16*(x+y*Data->iEdgedWidth), Data->iEdgedWidth);
-		bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, Data->scan_table, Data->lambda[i], Data->mpeg_quant_matrices);
+		bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i,
+								Data->scan_table, Data->lambda[i], Data->mpeg_quant_matrices, Data->quant_sq);
 		if (bits >= Data->iMinSAD[0]) return bits;
 	}
 
-	bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
+	bits += BITS_MULT * (xvid_cbpy_tab[15-(cbp>>2)].len - 2);
 
 	/*chroma U */
 	transfer_8to16subro(in, Data->CurU, vGMC->u + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2);
-	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, Data->scan_table, Data->lambda[4], Data->mpeg_quant_matrices);
+	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4,
+							Data->scan_table, Data->lambda[4], Data->mpeg_quant_matrices, Data->quant_sq);
 
 	if (bits >= Data->iMinSAD[0]) return bits;
 
 	/* chroma V */
 	transfer_8to16subro(in, Data->CurV , vGMC->v + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2);
-	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, Data->scan_table, Data->lambda[5], Data->mpeg_quant_matrices);
+	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5,
+							Data->scan_table, Data->lambda[5], Data->mpeg_quant_matrices, Data->quant_sq);
 
-	bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
+	bits += BITS_MULT * (mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len - 1);
 
 	*Data->cbp = cbp;
 
@@ -613,10 +639,10 @@
 	int mcsel = 0;
 	int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
 	const uint32_t iQuant = pMB->quant;
-
 	int min_rd, intra_rd, i, cbp;
 	VECTOR backup[5], *v;
 	Data->iQuant = iQuant;
+	Data->quant_sq = iQuant*iQuant;
 	Data->scan_table = VopFlags & XVID_VOP_ALTERNATESCAN ?
 						scan_tables[2] : scan_tables[0];
 
@@ -657,12 +683,15 @@
 			cbp = *Data->cbp;
 		}
 	}
-
-	intra_rd = findRD_intra(Data, pMB, x, y, pParam->mb_width);
-	if (intra_rd < min_rd) {
-		*Data->iMinSAD = min_rd = intra_rd;
-		mode = MODE_INTRA;
-		cbp = *Data->cbp;
+	
+	/* there is no way for INTRA to take less than 24 bits - go to findRD_intra() for calculations */
+	if (min_rd > 24*BITS_MULT) { 
+		intra_rd = findRD_intra(Data, pMB, x, y, pParam->mb_width);
+		if (intra_rd < min_rd) {
+			*Data->iMinSAD = min_rd = intra_rd;
+			mode = MODE_INTRA;
+			cbp = *Data->cbp;
+		}
 	}
 
 	pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0;
@@ -729,6 +758,8 @@
 						scan_tables[2] : scan_tables[0];
 
 	pMB->mcsel = 0;
+	Data->iQuant = iQuant;
+	Data->quant_sq = iQuant*iQuant;
 
 	for (i = 0; i < 6; i++) {
 		/* HVS models, anyone ? */