Rename constant name EMPTY to SLU_EMPTY in util_dist.h, to avoid name…

… clash with system defs.

EXAMPLE/Makefile

@@ -31,8 +31,8 @@
 include ../make.inc
 
 SEXM	= psdrive.o screate_matrix.o dcreate_matrix.o screate_A_x_b.o \
-	  psgssvx_tracking.o psgsrfs_tracking.o \
-	psutil.o sdistribute.o
+	  psgssvx_tracking.o psgsrfs_tracking.o 
+	#	psutil.o sdistribute.o
 # psgssvx_d2.o psgsrfs_d2.o sutil_dist.o psgsequb.o
 
 DEXM	= pddrive.o dcreate_matrix.o 

EXAMPLE/pddrive.c

@@ -117,6 +117,8 @@ int main(int argc, char *argv[])
      */
     set_default_options_dist(&options);
 #if 0
+    options.ParSymbFact       = YES;
+    options.ColPerm           = PARMETIS;
     options.RowPerm = LargeDiag_HWPM;
     options.IterRefine = NOREFINE;
     options.ColPerm = NATURAL;

SRC/communication_aux.c

@@ -40,7 +40,7 @@ int_t Wait_LSend
     for (int_t pj = 0; pj < Pc; ++pj)
     {
         /* Wait for Isend to complete before using lsub/lusup. */
-        if (ToSendR[lk][pj] != EMPTY && pj != mycol)
+        if (ToSendR[lk][pj] != SLU_EMPTY && pj != mycol)
         {
             MPI_Wait (&send_req[pj], &status);
             MPI_Wait (&send_req[pj + Pc], &status);

SRC/dcommunication_aux.c

@@ -47,7 +47,7 @@ int_t dIBcast_LPanel
 
     for (int_t pj = 0; pj < Pc; ++pj)
     {
-        if (ToSendR[lk][pj] != EMPTY)
+        if (ToSendR[lk][pj] != SLU_EMPTY)
         {
 
 
@@ -89,7 +89,7 @@ int_t dBcast_LPanel
 
     for (int_t pj = 0; pj < Pc; ++pj)
     {
-        if (ToSendR[lk][pj] != EMPTY)
+        if (ToSendR[lk][pj] != SLU_EMPTY)
         {
 
 

SRC/ddistribute.c

@@ -384,7 +384,7 @@ ddistribute(superlu_dist_options_t *options,
 
 	mem_use += (float) k*sizeof(int_t*) + (j + nsupers)*iword;
 
-	for (i = 0; i < j; ++i) index1[i] = EMPTY;
+	for (i = 0; i < j; ++i) index1[i] = SLU_EMPTY;
 	for (i = 0,j = 0; i < k; ++i, j += grid->npcol) ToSendR[i] = &index1[j];
 	k = CEILING( nsupers, grid->nprow ); /* Number of local block rows */
 
@@ -610,14 +610,14 @@ ddistribute(superlu_dist_options_t *options,
 	len = k * grid->nprow;
 	if ( !(index1 = int32Malloc_dist(len)) )
 	    ABORT("Malloc fails for fsendx_plist[0]");
-	for (i = 0; i < len; ++i) index1[i] = EMPTY;
+	for (i = 0; i < len; ++i) index1[i] = SLU_EMPTY;
 	for (i = 0, j = 0; i < k; ++i, j += grid->nprow)
 	    fsendx_plist[i] = &index1[j];
 	if ( !(bsendx_plist = (int **) SUPERLU_MALLOC(k*sizeof(int*))) )
 	    ABORT("Malloc fails for bsendx_plist[].");
 	if ( !(index1 = int32Malloc_dist(len)) )
 	    ABORT("Malloc fails for bsendx_plist[0]");
-	for (i = 0; i < len; ++i) index1[i] = EMPTY;
+	for (i = 0; i < len; ++i) index1[i] = SLU_EMPTY;
 	for (i = 0, j = 0; i < k; ++i, j += grid->nprow)
 	    bsendx_plist[i] = &index1[j];
 
@@ -675,7 +675,7 @@ ddistribute(superlu_dist_options_t *options,
 			pr = PROW( gb, grid );
 			if ( pr != jbrow &&
 			     myrow == jbrow &&  /* diag. proc. owning jb */
-			     bsendx_plist[ljb][pr] == EMPTY ) {
+			     bsendx_plist[ljb][pr] == SLU_EMPTY ) {
 			    bsendx_plist[ljb][pr] = YES;
 			    ++nbsendx;
                         }
@@ -742,7 +742,7 @@ ddistribute(superlu_dist_options_t *options,
 		    pr = PROW( gb, grid ); /* Process row owning this block */
 		    if ( pr != jbrow &&
 			 myrow == jbrow &&  /* diag. proc. owning jb */
-			 fsendx_plist[ljb][pr] == EMPTY /* first time */ ) {
+			 fsendx_plist[ljb][pr] == SLU_EMPTY /* first time */ ) {
 			fsendx_plist[ljb][pr] = YES;
 			++nfsendx;
                     }
@@ -1313,7 +1313,7 @@ ddistribute(superlu_dist_options_t *options,
 				if(Root==myrow){
 					rank_cnt_ref=1;
 					for (j = 0; j < grid->nprow; ++j) {
-						if ( fsendx_plist[ljb][j] != EMPTY ) {
+						if ( fsendx_plist[ljb][j] != SLU_EMPTY ) {
 							++rank_cnt_ref;
 						}
 					}
@@ -1651,7 +1651,7 @@ if ( !iam) printf(".. Construct Reduce tree for L: %.2f\t\n", t);
 				for (j = 0; j < grid->nprow; ++j) {
 					// printf("ljb %5d j %5d nprow %5d\n",ljb,j,grid->nprow);
 					// fflush(stdout);
-					if ( bsendx_plist[ljb][j] != EMPTY ) {
+					if ( bsendx_plist[ljb][j] != SLU_EMPTY ) {
 						++rank_cnt_ref;
 					}
 				}

SRC/dlook_ahead_update.c

@@ -250,7 +250,7 @@ while (j < nub && perm_u[2 * j] <= k0 + num_look_aheads)
 
         scp = &grid->rscp;      /* The scope of process row. */
         for (pj = 0; pj < Pc; ++pj) {
-            if (ToSendR[lk][pj] != EMPTY) {
+            if (ToSendR[lk][pj] != SLU_EMPTY) {
 #if ( PROFlevel>=1 )
                 TIC (t1);
 #endif
@@ -271,7 +271,7 @@ while (j < nub && perm_u[2 * j] <= k0 + num_look_aheads)
                         iam, kk, msgcnt[0], msgcnt[1], pj,
 			SLU_MPI_TAG(0,kk0), SLU_MPI_TAG(1,kk0));
 #endif
-            }  /* end if ( ToSendR[lk][pj] != EMPTY ) */
+            }  /* end if ( ToSendR[lk][pj] != SLU_EMPTY ) */
         } /* end for pj ... */
     } /* end if( look_ahead[kk] == k0 && kcol == mycol ) */
 } /* end while j < nub and perm_u[j] <k0+NUM_LOOK_AHEAD */

SRC/dnrformat_loc3d.c

@@ -68,7 +68,7 @@ void dGatherNRformat_loc3d
     if ( Fact == DOFACT ) { /* Factorize from scratch */
 	/* A3d is output. Compute counts from scratch */
 	A3d = SUPERLU_MALLOC(sizeof(NRformat_loc3d));
-	A3d->num_procs_to_send = EMPTY; // No X(2d) -> X(3d) comm. schedule yet
+	A3d->num_procs_to_send = SLU_EMPTY; // No X(2d) -> X(3d) comm. schedule yet
 	A2d = SUPERLU_MALLOC(sizeof(NRformat_loc));
 
 	// find number of nnzs
@@ -364,11 +364,11 @@ int dScatter_B3d(NRformat_loc3d *A3d,  // modified
          In the most general case, block rows of B are not of even size, then the
 	 Layer 0 partition may overlap with 3D partition in an arbitrary manner.
 	 For example:
-	                  P0        P1        P2       P3
+	                  P0          P1         P2        P3
              X on grid-0: |___________|__________|_________|________|
 
 	     X on 3D:     |___|____|_____|____|__|______|_____|_____|
-	                  P0  P1   P2    P3   P4   P5     P6   P7  
+	                  P0  P1   P2    P3   P4 P5     P6    P7  
 	*/
 	MPI_Status recv_status;
 	int pxy = grid2d->nprow * grid2d->npcol;
@@ -412,14 +412,14 @@ int dScatter_B3d(NRformat_loc3d *A3d,  // modified
 	    for (p = 0; p < nprocs; ++p) {
 		x_send_counts[p] = 0;
 		x_recv_counts[p] = 0;
-		procs_to_send_list[p] = EMPTY; // (-1)
-		procs_recv_from_list[p] = EMPTY;
+		procs_to_send_list[p] = SLU_EMPTY; // (-1)
+		procs_recv_from_list[p] = SLU_EMPTY;
 	    }
 
 	    /* All procs participate */
 	    MPI_Allgather(&(A3d->m_loc), 1, MPI_INT, m_loc_3d_counts, 1,
 			  MPI_INT, grid3d->comm);
-
+	    
 	    /* Layer 0 set up sends info. The other layers have 0 send counts. */
 	    if (grid3d->zscp.Iam == 0) {
 		int x_fst_row = A2d->fst_row; // start from a layer 0 boundary
@@ -461,15 +461,15 @@ int dScatter_B3d(NRformat_loc3d *A3d,  // modified
 	       x_recv_counts matrix */
 	    MPI_Alltoall(x_send_counts, 1, MPI_INT, x_recv_counts, 1, MPI_INT,
 			 grid3d->comm);
-
+	    
 	    j = 0; // tracking number procs to receive from
 	    for (p = 0; p < nprocs; ++p) {
 		if (x_recv_counts[p]) {
 		    procs_recv_from_list[j] = p;
 		    recv_count_list[j] = x_recv_counts[p];
 		    src = p;  tag = iam;
 		    ++j;
-#if 0
+#if 0		    
 		    printf("RECV: src %d -> iam %d, x_recv_counts[p] %d, tag %d\n",
 			   src, iam, x_recv_counts[p], tag);
 		    fflush(stdout);

SRC/get_perm_c_parmetis.c

@@ -605,8 +605,8 @@ a_plus_at_CompRow_loc
      THIS COUNTS FOR TWO PASSES OF THE LOCAL ROWS OF A AND T.   
      ------------------------------------------------------------------ */
 
-  /* Reset marker to EMPTY */
-  for (i = 0; i < n; ++i) marker[i] = EMPTY;
+  /* Reset marker to SLU_EMPTY */
+  for (i = 0; i < n; ++i) marker[i] = SLU_EMPTY;
   /* save rdispls information */
   for (p = 0; p < nprocs_i; p++)
     sdispls[p] = rdispls[p];
@@ -654,8 +654,8 @@ a_plus_at_CompRow_loc
   if (apat_mem > apat_mem_max)
     apat_mem_max = apat_mem;
 
-  /* Reset marker to EMPTY */
-  for (i = 0; i < n; i++) marker[i] = EMPTY;
+  /* Reset marker to SLU_EMPTY */
+  for (i = 0; i < n; i++) marker[i] = SLU_EMPTY;
   /* restore rdispls information */
   for (p = 0; p < nprocs_i; p++)
     rdispls[p] = sdispls[p];

SRC/pddistribute.c

@@ -695,7 +695,7 @@ pddistribute(superlu_dist_options_t *options, int_t n, SuperMatrix *A,
 
 	mem_use += (float) k*sizeof(int_t*) + (j + nsupers)*iword;
 
-	for (i = 0; i < j; ++i) index1[i] = EMPTY;
+	for (i = 0; i < j; ++i) index1[i] = SLU_EMPTY;
 	for (i = 0,j = 0; i < k; ++i, j += grid->npcol) ToSendR[i] = &index1[j];
 	k = CEILING( nsupers, grid->nprow ); /* Number of local block rows */
 
@@ -924,14 +924,14 @@ pddistribute(superlu_dist_options_t *options, int_t n, SuperMatrix *A,
 	len = k * grid->nprow;
 	if ( !(index1 = int32Malloc_dist(len)) )
 	    ABORT("Malloc fails for fsendx_plist[0]");
-	for (i = 0; i < len; ++i) index1[i] = EMPTY;
+	for (i = 0; i < len; ++i) index1[i] = SLU_EMPTY;
 	for (i = 0, j = 0; i < k; ++i, j += grid->nprow)
 	    fsendx_plist[i] = &index1[j];
 	if ( !(bsendx_plist = (int **) SUPERLU_MALLOC(k*sizeof(int*))) )
 	    ABORT("Malloc fails for bsendx_plist[].");
 	if ( !(index1 = int32Malloc_dist(len)) )
 	    ABORT("Malloc fails for bsendx_plist[0]");
-	for (i = 0; i < len; ++i) index1[i] = EMPTY;
+	for (i = 0; i < len; ++i) index1[i] = SLU_EMPTY;
 	for (i = 0, j = 0; i < k; ++i, j += grid->nprow)
 	    bsendx_plist[i] = &index1[j];
 	/* -------------------------------------------------------------- */
@@ -990,7 +990,7 @@ pddistribute(superlu_dist_options_t *options, int_t n, SuperMatrix *A,
 			pr = PROW( gb, grid );
 			if ( pr != jbrow &&
 			     myrow == jbrow &&  /* diag. proc. owning jb */
-			     bsendx_plist[ljb][pr] == EMPTY ) {
+			     bsendx_plist[ljb][pr] == SLU_EMPTY ) {
 			    bsendx_plist[ljb][pr] = YES;
 			    ++nbsendx;
                         }
@@ -1056,7 +1056,7 @@ pddistribute(superlu_dist_options_t *options, int_t n, SuperMatrix *A,
 		    pr = PROW( gb, grid ); /* Process row owning this block */
 		    if ( pr != jbrow &&
 			 myrow == jbrow &&  /* diag. proc. owning jb */
-			 fsendx_plist[ljb][pr] == EMPTY /* first time */ ) {
+			 fsendx_plist[ljb][pr] == SLU_EMPTY /* first time */ ) {
 			fsendx_plist[ljb][pr] = YES;
 			++nfsendx;
                     }
@@ -1630,7 +1630,7 @@ pddistribute(superlu_dist_options_t *options, int_t n, SuperMatrix *A,
 				if(Root==myrow){
 					rank_cnt_ref=1;
 					for (j = 0; j < grid->nprow; ++j) {
-						if ( fsendx_plist[ljb][j] != EMPTY ) {
+						if ( fsendx_plist[ljb][j] != SLU_EMPTY ) {
 							++rank_cnt_ref;
 						}
 					}
@@ -1974,7 +1974,7 @@ if ( !iam) printf(".. Construct Reduce tree for L: %.2f\t\n", t);
 				for (j = 0; j < grid->nprow; ++j) {
 					// printf("ljb %5d j %5d nprow %5d\n",ljb,j,grid->nprow);
 					// fflush(stdout);
-					if ( bsendx_plist[ljb][j] != EMPTY ) {
+					if ( bsendx_plist[ljb][j] != SLU_EMPTY ) {
 						++rank_cnt_ref;
 					}
 				}

SRC/pdgsmv.c

@@ -127,11 +127,11 @@ void pdgsmv_init
     }
 
     /* Build SPA to aid global to local translation. */
-    for (i = 0; i < n; ++i) spa[i] = EMPTY;
+    for (i = 0; i < n; ++i) spa[i] = SLU_EMPTY;
     for (i = 0; i < m_loc; ++i) { /* Loop through each row of A */
         for (j = rowptr[i]; j < rowptr[i+1]; ++j) {
 	    jcol = colind[j];
-	    if ( spa[jcol] == EMPTY ) { /* First time see this index */
+	    if ( spa[jcol] == SLU_EMPTY ) { /* First time see this index */
 	        p = row_to_proc[jcol];
 		if ( p == iam ) { /* Local */
 		  /*assert(jcol>=fst_row);*/

SRC/pdgstrf.c

@@ -1030,7 +1030,7 @@ pdgstrf(superlu_dist_options_t * options, int m, int n, double anorm,
         }
 
         for (pj = 0; pj < Pc; ++pj) {
-            if (ToSendR[lk][pj] != EMPTY) {
+            if (ToSendR[lk][pj] != SLU_EMPTY) {
 #if ( PROFlevel>=1 )
                 TIC (t1);
 #endif
@@ -1155,7 +1155,7 @@ pdgstrf(superlu_dist_options_t * options, int m, int n, double anorm,
                     }
                     scp = &grid->rscp;  /* The scope of process row. */
                     for (pj = 0; pj < Pc; ++pj) {
-                        if (ToSendR[lk][pj] != EMPTY) {
+                        if (ToSendR[lk][pj] != SLU_EMPTY) {
                             lusup1 = Lnzval_bc_ptr[lk];
 #if ( PROFlevel>=1 )
 			    TIC (t1);
@@ -1377,7 +1377,7 @@ pdgstrf(superlu_dist_options_t * options, int m, int n, double anorm,
 #endif
             for (pj = 0; pj < Pc; ++pj) {
                 /* Wait for Isend to complete before using lsub/lusup buffer. */
-                if (ToSendR[lk][pj] != EMPTY) {
+                if (ToSendR[lk][pj] != SLU_EMPTY) {
                     MPI_Wait (&send_req[pj], &status);
                     MPI_Wait (&send_req[pj + Pc], &status);
                 }
@@ -1704,7 +1704,7 @@ pdgstrf(superlu_dist_options_t * options, int m, int n, double anorm,
 
                         scp = &grid->rscp;  /* The scope of process row. */
                         for (pj = 0; pj < Pc; ++pj) {
-                            if (ToSendR[lk][pj] != EMPTY) {
+                            if (ToSendR[lk][pj] != SLU_EMPTY) {
 #if ( PROFlevel>=1 )
 			       TIC (t1);
 #endif

SRC/pdgstrs1.c

@@ -338,7 +338,7 @@ void pdgstrs1(superlu_dist_options_t *options, int_t n,
 		 * Send Xk to process column Pc[k].
 		 */
 		for (p = 0; p < Pr; ++p)
-		    if ( fsendx_plist[lk][p] != EMPTY ) {
+		    if ( fsendx_plist[lk][p] != SLU_EMPTY ) {
 			pi = PNUM( p, kcol, grid );
 #ifdef ISEND_IRECV
 			MPI_Isend( &x[ii - XK_H], knsupc * nrhs + XK_H,
@@ -454,7 +454,7 @@ void pdgstrs1(superlu_dist_options_t *options, int_t n,
 		   */
 		  kcol = PCOL( k, grid );
 		  for (p = 0; p < Pr; ++p)
-		      if ( fsendx_plist[lk][p] != EMPTY ) {
+		      if ( fsendx_plist[lk][p] != SLU_EMPTY ) {
 			  pi = PNUM( p, kcol, grid );
 #ifdef ISEND_IRECV
 			  MPI_Isend( &x[ii - XK_H], knsupc * nrhs + XK_H,
@@ -764,7 +764,7 @@ void pdgstrs1(superlu_dist_options_t *options, int_t n,
 		 * Send Xk to process column Pc[k].
 		 */
 		for (p = 0; p < Pr; ++p)
-		    if ( bsendx_plist[lk][p] != EMPTY ) {
+		    if ( bsendx_plist[lk][p] != SLU_EMPTY ) {
 			pi = PNUM( p, kcol, grid );
 #ifdef ISEND_IRECV
 			MPI_Isend( &x[ii - XK_H], knsupc * nrhs + XK_H,
@@ -855,7 +855,7 @@ void pdgstrs1(superlu_dist_options_t *options, int_t n,
 		     */
 		    kcol = PCOL( k, grid );
 		    for (p = 0; p < Pr; ++p)
-			if ( bsendx_plist[lk][p] != EMPTY ) {
+			if ( bsendx_plist[lk][p] != SLU_EMPTY ) {
 			    pi = PNUM( p, kcol, grid );
 #ifdef ISEND_IRECV
 			    MPI_Isend( &x[ii - XK_H], knsupc * nrhs + XK_H,

SRC/pdgstrs_Bglobal.c

@@ -362,7 +362,7 @@ pdgstrs_Bglobal(superlu_dist_options_t *options, int_t n,
 		 * Send Xk to process column Pc[k].
 		 */
 		for (p = 0; p < Pr; ++p) {
-		    if ( fsendx_plist[lk][p] != EMPTY ) {
+		    if ( fsendx_plist[lk][p] != SLU_EMPTY ) {
 			pi = PNUM( p, kcol, grid );
 #ifdef ISEND_IRECV
 			MPI_Isend( &x[ii - XK_H], knsupc * nrhs + XK_H,
@@ -487,7 +487,7 @@ pdgstrs_Bglobal(superlu_dist_options_t *options, int_t n,
 		   */
 		  kcol = PCOL( k, grid );
 		  for (p = 0; p < Pr; ++p) {
-		      if ( fsendx_plist[lk][p] != EMPTY ) {
+		      if ( fsendx_plist[lk][p] != SLU_EMPTY ) {
 			  pi = PNUM( p, kcol, grid );
 #ifdef ISEND_IRECV
 			  MPI_Isend( &x[ii - XK_H], knsupc * nrhs + XK_H,
@@ -798,7 +798,7 @@ pdgstrs_Bglobal(superlu_dist_options_t *options, int_t n,
 		 * Send Xk to process column Pc[k].
 		 */
 		for (p = 0; p < Pr; ++p) {
-		    if ( bsendx_plist[lk][p] != EMPTY ) {
+		    if ( bsendx_plist[lk][p] != SLU_EMPTY ) {
 			pi = PNUM( p, kcol, grid );
 #ifdef ISEND_IRECV
 			MPI_Isend( &x[ii - XK_H], knsupc * nrhs + XK_H,
@@ -895,7 +895,7 @@ pdgstrs_Bglobal(superlu_dist_options_t *options, int_t n,
 		     */
 		    kcol = PCOL( k, grid );
 		    for (p = 0; p < Pr; ++p) {
-			if ( bsendx_plist[lk][p] != EMPTY ) {
+			if ( bsendx_plist[lk][p] != SLU_EMPTY ) {
 			    pi = PNUM( p, kcol, grid );
 #ifdef ISEND_IRECV
 			    MPI_Isend( &x[ii - XK_H], knsupc * nrhs + XK_H,