===================================================================
RCS file: /home/cvs/OpenXM_contrib2/asir2000/engine/nd.c,v
retrieving revision 1.94
retrieving revision 1.97
diff -u -p -r1.94 -r1.97
--- OpenXM_contrib2/asir2000/engine/nd.c	2004/03/15 07:30:44	1.94
+++ OpenXM_contrib2/asir2000/engine/nd.c	2004/03/25 01:31:03	1.97
@@ -1,4 +1,4 @@
-/* $OpenXM: OpenXM_contrib2/asir2000/engine/nd.c,v 1.93 2004/03/13 07:48:30 noro Exp $ */
+/* $OpenXM: OpenXM_contrib2/asir2000/engine/nd.c,v 1.96 2004/03/17 08:16:24 noro Exp $ */
 
 #include "nd.h"
 
@@ -29,9 +29,12 @@ static NDV *nd_ps;
 static NDV *nd_ps_trace;
 static RHist *nd_psh;
 static int nd_psn,nd_pslen;
-
 static RHist *nd_red;
-
+static int *nd_work_vector;
+static int **nd_matrix;
+static int nd_matrix_len;
+static struct weight_or_block *nd_worb;
+static int nd_worb_len;
 static int nd_found,nd_create,nd_notfirst;
 static int nmv_adv;
 static int nd_demand;
@@ -422,6 +425,84 @@ int ndl_block_compare(UINT *d1,UINT *d2)
 	return 0;
 }
 
+int ndl_matrix_compare(UINT *d1,UINT *d2)
+{
+	int i,j,s;
+	int *v;
+
+	for ( j = 0; j < nd_nvar; j++ )
+		nd_work_vector[j] = GET_EXP(d1,j)-GET_EXP(d2,j);
+	for ( i = 0; i < nd_matrix_len; i++ ) {
+		v = nd_matrix[i];
+		for ( j = 0, s = 0; j < nd_nvar; j++ )
+			s += v[j]*nd_work_vector[j];
+		if ( s > 0 ) return 1;
+		else if ( s < 0 ) return -1;
+	}
+	return 0;
+}
+
+int ndl_composite_compare(UINT *d1,UINT *d2)
+{
+	int i,j,s,start,end,len,o;
+	int *v;
+	struct sparse_weight *sw;
+
+	for ( j = 0; j < nd_nvar; j++ )
+		nd_work_vector[j] = GET_EXP(d1,j)-GET_EXP(d2,j);
+	for ( i = 0; i < nd_worb_len; i++ ) {
+		len = nd_worb[i].length;
+		switch ( nd_worb[i].type ) {
+			case IS_DENSE_WEIGHT:
+				v = nd_worb[i].body.dense_weight;
+				for ( j = 0, s = 0; j < len; j++ )
+					s += v[j]*nd_work_vector[j];
+				if ( s > 0 ) return 1;
+				else if ( s < 0 ) return -1;
+				break;
+			case IS_SPARSE_WEIGHT:
+				sw = nd_worb[i].body.sparse_weight;
+				for ( j = 0, s = 0; j < len; j++ )
+					s += sw[j].value*nd_work_vector[sw[j].pos];
+				if ( s > 0 ) return 1;
+				else if ( s < 0 ) return -1;
+				break;
+			case IS_BLOCK:
+				o = nd_worb[i].body.block.order;
+				start = nd_worb[i].body.block.start;
+				switch ( o ) {
+					case 0:
+						end = start+len;
+						for ( j = start, s = 0; j < end; j++ )
+							s += MUL_WEIGHT(nd_work_vector[j],j);
+						if ( s > 0 ) return 1;
+						else if ( s < 0 ) return -1;
+						for ( j = end-1; j >= start; j-- )
+							if ( nd_work_vector[j] < 0 ) return 1;
+							else if ( nd_work_vector[j] > 0 ) return -1;
+						break;
+					case 1:
+						end = start+len;
+						for ( j = start, s = 0; j < end; j++ )
+							s += MUL_WEIGHT(nd_work_vector[j],j);
+						if ( s > 0 ) return 1;
+						else if ( s < 0 ) return -1;
+						for ( j = start; j < end; j++ )
+							if ( nd_work_vector[j] > 0 ) return 1;
+							else if ( nd_work_vector[j] < 0 ) return -1;
+						break;
+					case 2:
+						for ( j = start; j < end; j++ )
+							if ( nd_work_vector[j] > 0 ) return 1;
+							else if ( nd_work_vector[j] < 0 ) return -1;
+						break;
+				}
+				break;
+		}
+	}
+	return 0;
+}
+
 /* TDH -> WW -> TD-> RL */
 
 int ndl_ww_lex_compare(UINT *d1,UINT *d2)
@@ -873,6 +954,79 @@ ND nd_add(int mod,ND p1,ND p2)
 	}
 }
 
+/* XXX on opteron, the inlined manipulation of destructive additon of
+ * two NM seems to make gcc optimizer get confused, so the part is
+ * done in a function.
+ */
+
+int nm_destructive_add_q(NM *m1,NM *m2,NM *mr0,NM *mr)
+{
+	NM s;
+	Q t;
+	int can;
+
+	addq(CQ(*m1),CQ(*m2),&t);
+	s = *m1; *m1 = NEXT(*m1);
+	if ( t ) {
+		can = 1; NEXTNM2(*mr0,*mr,s); CQ(*mr) = (t);
+	} else {
+		can = 2; FREENM(s);
+	}
+	s = *m2; *m2 = NEXT(*m2); FREENM(s);
+	return can;
+}
+
+ND nd_add_q(ND p1,ND p2)
+{
+	int n,c,can;
+	ND r;
+	NM m1,m2,mr0,mr,s;
+	Q t;
+
+	if ( !p1 ) return p2;
+	else if ( !p2 ) return p1;
+	else {
+		can = 0;
+		for ( n = NV(p1), m1 = BDY(p1), m2 = BDY(p2), mr0 = 0; m1 && m2; ) {
+			c = DL_COMPARE(DL(m1),DL(m2));
+			switch ( c ) {
+				case 0:
+#if defined(__x86_64__)
+					can += nm_destructive_add_q(&m1,&m2,&mr0,&mr);
+#else
+					addq(CQ(m1),CQ(m2),&t);
+					s = m1; m1 = NEXT(m1);
+					if ( t ) {
+						can++; NEXTNM2(mr0,mr,s); CQ(mr) = (t);
+					} else {
+						can += 2; FREENM(s);
+					}
+					s = m2; m2 = NEXT(m2); FREENM(s);
+#endif
+					break;
+				case 1:
+					s = m1; m1 = NEXT(m1); NEXTNM2(mr0,mr,s);
+					break;
+				case -1:
+					s = m2; m2 = NEXT(m2); NEXTNM2(mr0,mr,s);
+					break;
+			}
+		}
+		if ( !mr0 )
+			if ( m1 ) mr0 = m1;
+			else if ( m2 ) mr0 = m2;
+			else return 0;
+		else if ( m1 ) NEXT(mr) = m1;
+		else if ( m2 ) NEXT(mr) = m2;
+		else NEXT(mr) = 0;
+		BDY(p1) = mr0;
+		SG(p1) = MAX(SG(p1),SG(p2));
+		LEN(p1) = LEN(p1)+LEN(p2)-can;
+		FREEND(p2);
+		return p1;
+	}
+}
+
 ND nd_add_sf(ND p1,ND p2)
 {
 	int n,c,can;
@@ -2551,14 +2705,14 @@ UINT *ndv_compute_bound(NDV p)
 int nd_get_exporigin(struct order_spec *ord)
 {
 	switch ( ord->id ) {
-		case 0:
+		case 0: case 2:
 			return 1;
 		case 1:
 			/* block order */
 			/* d[0]:weight d[1]:w0,...,d[nd_exporigin-1]:w(n-1) */
 			return ord->ord.block.length+1;
-		case 2:
-			error("nd_get_exporigin : matrix order is not supported yet.");
+		case 3:
+			error("nd_get_exporigin : composite order is not supported yet.");
 	}
 }
 
@@ -2603,6 +2757,7 @@ void nd_setup_parameters(int nvar,int max) {
 	nmv_adv = ROUND_FOR_ALIGN(sizeof(struct oNMV)+(nd_wpd-1)*sizeof(UINT));
 	nd_epos = nd_create_epos(nd_ord);
 	nd_blockmask = nd_create_blockmask(nd_ord);
+	nd_work_vector = (int *)REALLOC(nd_work_vector,nd_nvar*sizeof(int));
 }
 
 ND_pairs nd_reconstruct(int mod,int trace,ND_pairs d)
@@ -3354,14 +3509,29 @@ void nd_init_ord(struct order_spec *ord)
 			}
 			break;
 		case 1:
+			/* block order */
 			/* XXX */
 			nd_dcomp = -1;
 			nd_isrlex = 0;
 			ndl_compare_function = ndl_block_compare;
 			break;
 		case 2:
-			error("nd_init_ord : matrix order is not supported yet.");
+			/* matrix order */
+			/* XXX */
+			nd_dcomp = -1;
+			nd_isrlex = 0;
+			nd_matrix_len = ord->ord.matrix.row;
+			nd_matrix = ord->ord.matrix.matrix;
+			ndl_compare_function = ndl_matrix_compare;
 			break;
+		case 3:
+			/* composite order */
+			nd_dcomp = -1;
+			nd_isrlex = 0;
+			nd_worb_len = ord->ord.composite.length;
+			nd_worb = ord->ord.composite.w_or_b;
+			ndl_compare_function = ndl_composite_compare;
+			break;
 	}
 	nd_ord = ord;
 }
@@ -3372,7 +3542,8 @@ BlockMask nd_create_blockmask(struct order_spec *ord)
 	UINT *t;
 	BlockMask bm;
 
-	if ( !ord->id )
+	/* we only create mask table for block order */
+	if ( ord->id != 1 )
 		return 0;
 	n = ord->ord.block.length;
 	bm = (BlockMask)MALLOC(sizeof(struct oBlockMask));
@@ -3430,7 +3601,14 @@ EPOS nd_create_epos(struct order_spec *ord)
 			}
 			break;
 		case 2:
-			error("nd_create_epos : matrix order is not supported yet.");
+			/* matrix order */
+		case 3:
+			/* composite order */
+			for ( i = 0; i < nd_nvar; i++ ) {
+				epos[i].i = nd_exporigin + i/nd_epw;
+				epos[i].s = (nd_epw-(i%nd_epw)-1)*nd_bpe;
+			}
+			break;
 	}
 	return epos;
 }