Merge branch 'serial-match' into development

2026-10-06 14:44:55 +00:00 · 2025-03-21 14:10:39 +01:00
parent 9ad88b1355 e922582aad
commit b9cf9dca06
51 changed files with 388 additions and 208 deletions
@@ -73,6 +73,22 @@ module amg_d_matchboxp_mod
  use iso_c_binding
  use psb_base_cbind_mod

+#if defined(PSB_SERIAL_MPI)
+  interface MatchingC
+    subroutine dMatchingC(nlver,nledge,verlocptr,verlocind,edgelocweight,&
+         & verdistance, mate) bind(c,name='dMatching')
+      use iso_c_binding
+      import :: psb_c_ipk_, psb_c_lpk_
+      implicit none
+
+      integer(psb_c_lpk_), value :: nlver,nledge
+      integer(psb_c_lpk_) :: verlocptr(*),verlocind(*), verdistance(*)
+      integer(psb_c_lpk_) :: mate(*)
+      real(c_double) :: edgelocweight(*)
+    end subroutine dMatchingC
+  end interface MatchingC
+
+#else
  interface MatchBoxPC
    subroutine dMatchBoxPC(nlver,nledge,verlocptr,verlocind,edgelocweight,&
         & verdistance, mate, myrank, numprocs, icomm,&
@@ -93,7 +109,7 @@ module amg_d_matchboxp_mod
      real(c_double)     :: msgpercent(*)
    end subroutine dMatchBoxPC
  end interface MatchBoxPC
-
+#endif
  interface amg_i_aggr_assign
    module procedure amg_i_d_aggr_assign
  end interface amg_i_aggr_assign
@@ -1129,11 +1145,15 @@ contains
      call psb_barrier(ictxt)
      if (me == 0) write(0,*)' Calling MatchBoxP '
    end if
-
+#if defined(PSB_SERIAL_MPI)
+    call MatchingC(nlver,nledge,verlocptr,verlocind,edgelocweight,&
+         & verdistance, mate)
+#else
    call MatchBoxPC(nlver,nledge,verlocptr,verlocind,edgelocweight,&
         & verdistance, mate, mrank, mnp, icomm,&
         & msgindsent,msgactualsent,msgpercent,&
         & ph0_time, ph1_time, ph2_time, ph1_card, ph2_card)
+#endif
    verlocptr(:)   = verlocptr(:)  + 1
    verlocind(:)   = verlocind(:) + 1
    verdistance(:) = verdistance(:) + 1
@@ -119,10 +119,6 @@
 module amg_d_parmatch_aggregator_mod
  use amg_d_base_aggregator_mod
  use amg_d_matchboxp_mod
-#if defined(PSB_SERIAL_MPI)
-  type, extends(amg_d_base_aggregator_type) :: amg_d_parmatch_aggregator_type
-  end type amg_d_parmatch_aggregator_type
-#else 
  type, extends(amg_d_base_aggregator_type) :: amg_d_parmatch_aggregator_type
    integer(psb_ipk_) :: matching_alg
    integer(psb_ipk_) :: n_sweeps   ! When n_sweeps >1 we need an auxiliary descriptor
@@ -685,5 +681,4 @@ contains

    return
  end subroutine amg_d_parmatch_aggregator_bld_map
-#endif
 end module amg_d_parmatch_aggregator_mod
@@ -73,6 +73,22 @@ module amg_s_matchboxp_mod
  use iso_c_binding
  use psb_base_cbind_mod

+#if defined(PSB_SERIAL_MPI)
+  interface MatchingC
+    subroutine sMatchingC(nlver,nledge,verlocptr,verlocind,edgelocweight,&
+         & verdistance, mate) bind(c,name='sMatching')
+      use iso_c_binding
+      import :: psb_c_ipk_, psb_c_lpk_
+      implicit none
+
+      integer(psb_c_lpk_), value :: nlver,nledge
+      integer(psb_c_lpk_) :: verlocptr(*),verlocind(*), verdistance(*)
+      integer(psb_c_lpk_) :: mate(*)
+      real(c_float) :: edgelocweight(*)
+    end subroutine sMatchingC
+  end interface MatchingC
+
+#else
  interface MatchBoxPC
    subroutine sMatchBoxPC(nlver,nledge,verlocptr,verlocind,edgelocweight,&
         & verdistance, mate, myrank, numprocs, icomm,&
@@ -93,7 +109,7 @@ module amg_s_matchboxp_mod
      real(c_double)     :: msgpercent(*)
    end subroutine sMatchBoxPC
  end interface MatchBoxPC
-
+#endif
  interface amg_i_aggr_assign
    module procedure amg_i_s_aggr_assign
  end interface amg_i_aggr_assign
@@ -1129,11 +1145,15 @@ contains
      call psb_barrier(ictxt)
      if (me == 0) write(0,*)' Calling MatchBoxP '
    end if
-
+#if defined(PSB_SERIAL_MPI)
+    call MatchingC(nlver,nledge,verlocptr,verlocind,edgelocweight,&
+         & verdistance, mate)
+#else
    call MatchBoxPC(nlver,nledge,verlocptr,verlocind,edgelocweight,&
         & verdistance, mate, mrank, mnp, icomm,&
         & msgindsent,msgactualsent,msgpercent,&
         & ph0_time, ph1_time, ph2_time, ph1_card, ph2_card)
+#endif
    verlocptr(:)   = verlocptr(:)  + 1
    verlocind(:)   = verlocind(:) + 1
    verdistance(:) = verdistance(:) + 1
@@ -119,10 +119,6 @@
 module amg_s_parmatch_aggregator_mod
  use amg_s_base_aggregator_mod
  use amg_s_matchboxp_mod
-#if defined(PSB_SERIAL_MPI)
-  type, extends(amg_s_base_aggregator_type) :: amg_s_parmatch_aggregator_type
-  end type amg_s_parmatch_aggregator_type
-#else 
  type, extends(amg_s_base_aggregator_type) :: amg_s_parmatch_aggregator_type
    integer(psb_ipk_) :: matching_alg
    integer(psb_ipk_) :: n_sweeps   ! When n_sweeps >1 we need an auxiliary descriptor
@@ -685,5 +681,4 @@ contains

    return
  end subroutine amg_s_parmatch_aggregator_bld_map
-#endif
 end module amg_s_parmatch_aggregator_mod
@@ -5,7 +5,8 @@ MODDIR=../../../modules
 HERE=../..

 FINCLUDES=$(FMFLAG)$(HERE) $(FMFLAG)$(MODDIR) $(FMFLAG)$(INCDIR) $(PSBLAS_INCLUDES)
-CXXINCLUDES=$(FMFLAG)$(HERE) $(FMFLAG)$(INCDIR) $(FMFLAG)/.
+CXXINCLUDES=$(FMFLAG)$(HERE) $(FMFLAG)$(INCDIR) $(FMFLAG)/. -I$(PSBLAS_INCDIR)
+CINCLUDES=$(FMFLAG)$(HERE) $(FMFLAG)$(INCDIR) $(FMFLAG)/. -I$(PSBLAS_INCDIR)

 #CINCLUDES= -I${SUPERLU_INCDIR} -I${HSL_INCDIR}  -I${SPRAL_INCDIR} -I/home/users/pasqua/Ambra/BootCMatch/include -lBCM -L/home/users/pasqua/Ambra/BootCMatch/lib  -lm

@@ -61,7 +62,8 @@ amg_s_parmatch_unsmth_bld.o \
 amg_s_parmatch_smth_bld.o \
 amg_s_parmatch_spmm_bld_inner.o

-MPCOBJS=MatchBoxPC.o \
+MPCOBJS=Matching.o \
+MatchBoxPC.o \
 sendBundledMessages.o \
 initialize.o \
 extractUChunk.o \
@@ -40,15 +40,17 @@
 // ************************************************************************
 #include <stdio.h>
 #include <stdlib.h>
-#if !defined(SERIAL_MPI)
+#include "amg_config.h"
+#include "MatchBoxPC.h"
+#if !defined(PSB_SERIAL_MPI)
 #include <mpi.h>
 #endif

-#include "MatchBoxPC.h"
 #ifdef __cplusplus
 extern "C" {
 #endif

+#if !defined(PSB_SERIAL_MPI)

 void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
 		MilanLongInt* verLocPtr, MilanLongInt* verLocInd, MilanReal* edgeLocWeight,
@@ -58,7 +60,6 @@ void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
 		MilanLongInt* msgIndSent, MilanLongInt* msgActualSent, MilanReal* msgPercent,
 		MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
 		MilanLongInt* ph1_card, MilanLongInt* ph2_card ) {
-#if !defined(SERIAL_MPI)
  MPI_Comm C_comm=MPI_Comm_f2c(icomm);

 #ifdef DEBUG
@@ -72,7 +73,7 @@ void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
  double tmr = MPI_Wtime();
 #endif

-#if defined(OPENMP)
+#if defined(PSB_OPENMP)
 	//fprintf(stderr,"Warning: using buggy OpenMP matching!\n");
        dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(NLVer, NLEdge,
 							   verLocPtr, verLocInd, edgeLocWeight,
@@ -97,7 +98,6 @@ void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
 	fprintf(stderr, "Elaboration time: %f for %ld nodes\n", tmr, NLVer);
 #endif
 	
-#endif
 }

 void sMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
@@ -108,13 +108,12 @@ void sMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
 		MilanLongInt* msgIndSent, MilanLongInt* msgActualSent, MilanReal* msgPercent,
 		MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
 		MilanLongInt* ph1_card, MilanLongInt* ph2_card ) {
-#if !defined(SERIAL_MPI)
  MPI_Comm C_comm=MPI_Comm_f2c(icomm);
 #ifdef DEBUG
  fprintf(stderr,"MatchBoxPC: rank %d nlver %ld nledge %ld [ %ld %ld ]\n",
 	  myRank,NLVer, NLEdge,verDistance[0],verDistance[1]);
 #endif
-#if defined(OPENMP)
+#if defined(PSB_OPENMP)
 	//fprintf(stderr,"Warning: using buggy OpenMP matching!\n");
  salgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(NLVer, NLEdge,
 							      verLocPtr, verLocInd, edgeLocWeight,
@@ -132,9 +131,10 @@ void sMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
 							    ph0_time, ph1_time, ph2_time,
 							    ph1_card, ph2_card );
 #endif
-#endif
 }

+#endif
+  
 #ifdef __cplusplus
 }
 #endif
@@ -59,7 +59,10 @@
 #include <assert.h>
 #include <map>
 #include <vector>
-#ifdef OPENMP
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)
+
+#ifdef PSB_OPENMP
 // OpenMP is included and used if and only if the OpenMP version of the matching
 // is required
 #include "omp.h"
@@ -82,7 +85,6 @@ const int BundleTag = 9;  // Predefined tag

 static vector<MilanLongInt> DEFAULT_VECTOR;

-#if !defined(SERIAL_MPI)

 // MPI type map
 template <typename T>
@@ -95,14 +97,12 @@ template <>
 inline MPI_Datatype TypeMap<double>() { return MPI_DOUBLE; }
 template <>
 inline MPI_Datatype TypeMap<float>() { return MPI_FLOAT; }
-#endif

 #ifdef __cplusplus
 extern "C"
 {
 #endif

-#if !defined(SERIAL_MPI)

 #define MilanMpiLongInt MPI_LONG_LONG

@@ -630,8 +630,9 @@ is disabled there is no reason to actually compile or reference them. */
                     MilanReal *ph0_time, MilanReal *ph1_time, MilanReal *ph2_time,
                     MilanLongInt *ph1_card, MilanLongInt *ph2_card);

-#endif
+
 #ifdef __cplusplus
 }
 #endif
 #endif
+#endif
@@ -0,0 +1,227 @@
+#include "amg_config.h"
+#if defined(PSB_SERIAL_MPI)
+#include <stdlib.h>
+#include "psb_types.h"
+
+psb_l_t d_trymatch(psb_l_t rowindex, psb_l_t colindex,
+		     psb_l_t nrows_W, psb_l_t *W_i, psb_l_t *W_j, psb_d_t *W_data,
+		     psb_l_t *jrowindex, psb_l_t ljrowindex,
+		     psb_l_t *jcolindex, psb_l_t ljcolindex, psb_l_t *rmatch)
+{
+
+  psb_l_t tryrowmatch, trycolmatch;
+  psb_l_t i, j, k, nzrow_W, startj, kindex;
+  psb_d_t cweight, nweight; 
+
+  //  psb_l_t     *W_i = bcm_CSRMatrixI(W);
+  //  psb_l_t     *W_j = bcm_CSRMatrixJ(W);
+  //  psb_l_t     nrows_W  =  bcm_CSRMatrixNumRows(W);
+  //  psb_d_t  *W_data=bcm_CSRMatrixData(W);
+
+  k=-1;
+  i=0;
+  while (i<ljrowindex && k ==-1)
+    {
+      if(jrowindex[i]==colindex) k=i;
+      i++;
+    }
+  if(k >= 0)
+    {
+      ljrowindex=ljrowindex-1;
+      for(i=k; i<ljrowindex; ++i) jrowindex[i]=jrowindex[i+1];
+    }
+
+  k=-1;
+  i=0;
+  while(i<ljcolindex && k==-1)
+    {
+      if(jcolindex[i]==rowindex) k=i;
+      i++;
+    }
+  if(k >= 0)
+    {
+      ljcolindex=ljcolindex-1;
+      for(i=k; i<ljcolindex; ++i) jcolindex[i]=jcolindex[i+1];
+    }
+
+  nzrow_W=W_i[rowindex+1]-W_i[rowindex];
+  startj=W_i[rowindex];
+  kindex=-1;
+  k=0;
+  while(k<nzrow_W && kindex==-1) 
+    {
+      if(W_j[startj+k]==colindex) kindex=startj+k;
+      k++;
+    }
+  cweight=W_data[kindex];
+
+  while((rmatch[rowindex]==-1 && rmatch[colindex]==-1)
+	&& (ljrowindex !=0 || ljcolindex !=0))    
+    {
+
+      if (rmatch[rowindex] == -1 && ljrowindex !=0)
+	{
+	  tryrowmatch=jrowindex[0];
+	  nzrow_W=W_i[rowindex+1]-W_i[rowindex];
+	  startj=W_i[rowindex];
+	  k=0;
+	  kindex=-1;
+	  while(k<nzrow_W && kindex==-1) 
+	    {
+	      if(W_j[startj+k]==tryrowmatch) kindex=startj+k;
+	      k++;
+	    }
+	  nweight=W_data[kindex];
+
+	  ljrowindex=ljrowindex-1;
+	  for(i=0; i<ljrowindex; ++i) jrowindex[i]=jrowindex[i+1];
+        
+	  if(nweight > cweight && rmatch[tryrowmatch]==-1)
+	    {
+          
+              nzrow_W=W_i[tryrowmatch+1]-W_i[tryrowmatch];
+	      psb_l_t *trymatchindexrow;
+	      trymatchindexrow= (psb_l_t *) calloc(nzrow_W, sizeof(psb_l_t));
+
+	      startj=W_i[tryrowmatch];
+	      for(k=0; k<nzrow_W; ++k) trymatchindexrow[k]=W_j[startj+k]; 
+
+	      d_trymatch(rowindex,tryrowmatch,
+			   nrows_W, W_i, W_j, W_data,
+			   jrowindex,ljrowindex,
+			   trymatchindexrow,nzrow_W,rmatch);
+	      free(trymatchindexrow);
+	    }
+	}
+
+      if(rmatch[colindex]==-1 && ljcolindex!=0)
+	{
+	  trycolmatch=jcolindex[0];
+	  nzrow_W=W_i[colindex+1]-W_i[colindex];
+	  startj=W_i[colindex];
+	  k=0;
+	  kindex=-1;
+	  while(k<nzrow_W && kindex==-1) 
+	    {
+	      if(W_j[startj+k]==trycolmatch) kindex=startj+k;
+	      k++;
+	    }
+	  nweight=W_data[kindex];
+
+	  ljcolindex=ljcolindex-1;
+	  for(i=0; i<ljcolindex; ++i) jcolindex[i]=jcolindex[i+1];
+        
+	  if(nweight > cweight && rmatch[trycolmatch]==-1)
+	    {
+	      nzrow_W=W_i[trycolmatch+1]-W_i[trycolmatch];
+	      psb_l_t *trymatchindexcol;
+	      trymatchindexcol= (psb_l_t *) calloc(nzrow_W, sizeof(psb_l_t));
+
+	      startj=W_i[trycolmatch];
+	      for(k=0; k<nzrow_W; ++k) trymatchindexcol[k]=W_j[startj+k];
+
+	      d_trymatch(colindex,trycolmatch,
+			   nrows_W, W_i, W_j, W_data,			   
+			   jcolindex,ljcolindex,
+			   trymatchindexcol,nzrow_W,rmatch);
+	      free(trymatchindexcol);
+	    }
+	}
+    }
+        
+  if(rmatch[rowindex]==-1 & rmatch[colindex]==-1)
+    {
+      rmatch[rowindex]=colindex;
+      rmatch[colindex]=rowindex;
+    }
+
+  return 0;
+}
+
+
+
+
+psb_l_t *d_CSRMatrixHMatch( psb_l_t  nrows_B,  psb_l_t  ncols_B,
+			       psb_l_t  nnz_B,  psb_l_t     *B_i,
+			       psb_l_t     *B_j, psb_d_t  *B_data)
+{
+
+  psb_l_t i, j, k, *rmatch;
+  psb_d_t  *c, alpha;
+  psb_l_t jbp, nzrows_B;
+  psb_d_t tmp=0.0;
+  psb_l_t rno, cno, nzrows_cno, startj, ljrowindex, ljjrowindex;
+  psb_l_t *jrowindex, *jjrowindex;
+
+#if 0
+  psb_l_t     *B_i = bcm_CSRMatrixI(B);
+  psb_l_t     *B_j = bcm_CSRMatrixJ(B);
+  psb_d_t  *B_data = bcm_CSRMatrixData(B);
+  psb_l_t  nrows_B  =  bcm_CSRMatrixNumRows(B);
+  psb_l_t  ncols_B  =  bcm_CSRMatrixNumCols(B);
+  psb_l_t  nnz_B  =  bcm_CSRMatrixNumNonzeros(B);
+#endif
+  //  assert(nrows_B==ncols_B);
+
+  rmatch = (psb_l_t *) calloc(nrows_B, sizeof(psb_l_t));
+
+  for(i=0; i<nrows_B; ++i)  rmatch[i]=-1;
+
+  jrowindex = (psb_l_t *) calloc(nrows_B, sizeof(psb_l_t));
+  jjrowindex = (psb_l_t *) calloc(nrows_B, sizeof(psb_l_t));
+
+  jbp=0;
+  for(i=0; i< nrows_B; ++i)
+    { 
+      nzrows_B=B_i[i+1]-B_i[i];
+
+      for(j=0; j<nzrows_B; ++j) jrowindex[j]=B_j[jbp+j];
+      for(j=0; j<nzrows_B; ++j)
+	{
+	  rno=i;
+	  cno=B_j[jbp+j];
+
+	  startj=B_i[cno];
+	  nzrows_cno=B_i[cno+1]-startj;
+
+	  for(k=0; k<nzrows_cno; ++k) jjrowindex[k]=B_j[startj+k];
+
+	  if(rmatch[rno] == -1 && rmatch[cno] == -1)
+	    d_trymatch(rno,cno,nrows_B,B_i,B_j,B_data,
+			 jrowindex,nzrows_B,jjrowindex,nzrows_cno,rmatch);
+        
+	}
+
+      jbp=jbp+nzrows_B;
+    }
+   
+  free(jrowindex);
+  free(jjrowindex);
+  
+  return rmatch;
+}
+
+
+void dMatching(psb_l_t NLVer, psb_l_t NLEdge,
+	       psb_l_t *verLocPtr, psb_l_t *verLocInd, psb_d_t *edgeLocWeight,
+	       psb_l_t *verDistance, psb_l_t *Mate)
+{
+  psb_l_t *lmate;
+  lmate = d_CSRMatrixHMatch( NLVer, NLVer,   NLEdge,
+			       verLocPtr, verLocInd,edgeLocWeight);
+  for (psb_l_t i=0; i<NLVer; i++){
+    Mate[i] = lmate[i];
+  }
+  free(lmate);
+  return;
+}
+
+void sMatching(psb_l_t NLVer, psb_l_t NLEdge,
+	       psb_l_t *verLocPtr, psb_l_t *verLocInd, psb_s_t *edgeLocWeight,
+	       psb_l_t *verDistance, psb_l_t *Mate)
+{
+
+  return;
+}
+
+#endif
@@ -0,0 +1,11 @@
+#include "amg_config.h"
+
+#if defined(PSB_SERIAL_MPI)
+#include "psb_types.h"
+void dMatching(psb_l_t NLVer, psb_l_t NLEdge,
+	       psb_l_t *verLocPtr, psb_l_t *verLocInd, psb_d_t *edgeLocWeight,
+	       psb_l_t *verDistance, psb_l_t *Mate);
+void sMatching(psb_l_t NLVer, psb_l_t NLEdge,
+	       psb_l_t *verLocPtr, psb_l_t *verLocInd, psb_s_t *edgeLocWeight,
+	       psb_l_t *verDistance, psb_l_t *Mate);
+#endif
@@ -70,7 +70,7 @@
 Statistics: ph1_card, ph2_card : Size: |P| number of processes in the comm-world (number of matched edges in Phase 1 and Phase 2)
 */

-#ifdef SERIAL_MPI
+#ifdef PSB_SERIAL_MPI
 #else

 // DOUBLE PRECISION VERSION
@@ -86,7 +86,7 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
 							      MilanReal* msgPercent,
                                                              MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
                                                              MilanLongInt* ph1_card, MilanLongInt* ph2_card ) {
-#if !defined(SERIAL_MPI)
+#if !defined(PSB_SERIAL_MPI)
 #ifdef PRINT_DEBUG_INFO_
    cout<<"\n("<<myRank<<")Within algoEdgeApproxDominatingEdgesLinearSearchMessageBundling()"; fflush(stdout);
 #endif
@@ -1313,7 +1313,7 @@ void salgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
      MilanReal* msgPercent,
      MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
      MilanLongInt* ph1_card, MilanLongInt* ph2_card ) {
-#if !defined(SERIAL_MPI)
+#if !defined(PSB_SERIAL_MPI)
 #ifdef PRINT_DEBUG_INFO_
    cout<<"\n("<<myRank<<")Within algoEdgeApproxDominatingEdgesLinearSearchMessageBundling()"; fflush(stdout);
 #endif
@@ -70,7 +70,7 @@
 Statistics: ph1_card, ph2_card : Size: |P| number of processes in the comm-world (number of matched edges in Phase 1 and Phase 2)
 */
 //#define DEBUG_HANG_
-#ifdef SERIAL_MPI
+#ifdef PSB_SERIAL_MPI
 #else

 // DOUBLE PRECISION VERSION
@@ -102,7 +102,7 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
     *            i+1-th value is the position of the first element on the i+1-th row
     */

-#if !defined(SERIAL_MPI)
+#if !defined(PSB_SERIAL_MPI)
 #ifdef PRINT_DEBUG_INFO_
    cout << "\n(" << myRank << ")Within algoEdgeApproxDominatingEdgesLinearSearchMessageBundling()";
    fflush(stdout);
@@ -583,7 +583,7 @@ void salgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
     *            i+1-th value is the position of the first element on the i+1-th row
     */

-#if !defined(SERIAL_MPI)
+#if !defined(PSB_SERIAL_MPI)
 #ifdef PRINT_DEBUG_INFO_
    cout << "\n(" << myRank << ")Within algoEdgeApproxDominatingEdgesLinearSearchMessageBundling()";
    fflush(stdout);
@@ -98,11 +98,7 @@ subroutine  amg_d_parmatch_aggregator_inner_mat_asb(ag,parms,a,desc_a,&
     & ac,desc_ac, op_prol,op_restr,info)
  use psb_base_mod
  use amg_base_prec_type
-#if defined(PSB_SERIAL_MPI)
-    use amg_d_parmatch_aggregator_mod
-#else
  use amg_d_parmatch_aggregator_mod, amg_protect_name => amg_d_parmatch_aggregator_inner_mat_asb
-#endif
  implicit none
  class(amg_d_parmatch_aggregator_type), target, intent(inout) :: ag
  type(amg_dml_parms), intent(inout)    :: parms
@@ -133,8 +129,6 @@ subroutine  amg_d_parmatch_aggregator_inner_mat_asb(ag,parms,a,desc_a,&
  ictxt = desc_a%get_context()
  call psb_info(ictxt,me,np)

-#if !defined(PSB_SERIAL_MPI)
-
  if (debug) write(0,*) me,' ',trim(name),' Start:',&
       & allocated(ag%ac),allocated(ag%desc_ac), allocated(ag%prol),allocated(ag%restr)

@@ -146,16 +140,17 @@ subroutine  amg_d_parmatch_aggregator_inner_mat_asb(ag,parms,a,desc_a,&
  case(amg_repl_mat_)
    !
    !
-    info = psb_err_internal_error_
-    call psb_errpush(info,name,a_err='no repl coarse_mat_ here')
-    goto 9999
-
+    if (np>1) then 
+      info = psb_err_internal_error_
+      call psb_errpush(info,name,a_err='no repl coarse_mat_ here')
+      goto 9999
+    end if
  case default
    info = psb_err_internal_error_
    call psb_errpush(info,name,a_err='invalid amg_coarse_mat_')
    goto 9999
  end select
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -98,11 +98,7 @@ subroutine  amg_d_parmatch_aggregator_mat_asb(ag,parms,a,desc_a,&
     & ac,desc_ac, op_prol,op_restr,info)
  use psb_base_mod
  use amg_base_prec_type
-#if defined(PSB_SERIAL_MPI)
-    use amg_d_parmatch_aggregator_mod
-#else
  use amg_d_parmatch_aggregator_mod, amg_protect_name => amg_d_parmatch_aggregator_mat_asb
-#endif
  implicit none
  class(amg_d_parmatch_aggregator_type), target, intent(inout) :: ag
  type(amg_dml_parms), intent(inout)    :: parms
@@ -135,8 +131,6 @@ subroutine  amg_d_parmatch_aggregator_mat_asb(ag,parms,a,desc_a,&
    return
  end if

-
-#if !defined(PSB_SERIAL_MPI)
  if (debug) write(0,*) me,' ',trim(name),' Start:',&
       & allocated(ag%ac),allocated(ag%desc_ac), allocated(ag%prol),allocated(ag%restr)

@@ -199,7 +193,7 @@ subroutine  amg_d_parmatch_aggregator_mat_asb(ag,parms,a,desc_a,&
    call psb_errpush(info,name,a_err='invalid amg_coarse_mat_')
    goto 9999
  end select
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -135,11 +135,7 @@ subroutine  amg_d_parmatch_aggregator_mat_bld(ag,parms,a,desc_a,ilaggr,nlaggr,&
  use psb_base_mod
  use amg_d_inner_mod
  use amg_base_prec_type
-#if defined(PSB_SERIAL_MPI)
-    use amg_d_parmatch_aggregator_mod
-#else
  use amg_d_parmatch_aggregator_mod, amg_protect_name => amg_d_parmatch_aggregator_mat_bld
-#endif
  implicit none

  class(amg_d_parmatch_aggregator_type), target, intent(inout) :: ag
@@ -176,7 +172,6 @@ subroutine  amg_d_parmatch_aggregator_mat_bld(ag,parms,a,desc_a,ilaggr,nlaggr,&
  ! algorithm specified by
  !

-#if !defined(PSB_SERIAL_MPI)
  call clean_shortcuts(ag)
  !
  ! When requesting smoothed aggregation we cannot use the
@@ -212,14 +207,12 @@ subroutine  amg_d_parmatch_aggregator_mat_bld(ag,parms,a,desc_a,ilaggr,nlaggr,&
    call psb_errpush(psb_err_from_subroutine_,name,a_err='Inner aggrmat asb')
    goto 9999
  end if
-#endif
  call psb_erractionrestore(err_act)
  return

 9999 call psb_error_handler(err_act)
  return

-#if !defined(PSB_SERIAL_MPI)

 contains
  subroutine clean_shortcuts(ag)
@@ -248,5 +241,4 @@ contains
      end if
    end if
  end subroutine clean_shortcuts
-#endif
 end subroutine amg_d_parmatch_aggregator_mat_bld
@@ -47,11 +47,7 @@ subroutine  amg_d_parmatch_aggregator_build_tprol(ag,parms,ag_data,&
  use psb_base_mod
  use amg_base_prec_type
  use amg_d_inner_mod
-#if defined(PSB_SERIAL_MPI)
-  use amg_d_parmatch_aggregator_mod
-#else
  use amg_d_parmatch_aggregator_mod, amg_protect_name => amg_d_parmatch_aggregator_build_tprol
-#endif
  use iso_c_binding
  implicit none
  class(amg_d_parmatch_aggregator_type), target, intent(inout) :: ag
@@ -120,8 +116,6 @@ subroutine  amg_d_parmatch_aggregator_build_tprol(ag,parms,ag_data,&
       &   amg_aggr_ord_nat_,is_legal_ml_aggr_ord)
  call amg_check_def(parms%aggr_thresh,'Aggr_Thresh',dzero,is_legal_d_aggr_thrs)

-#if !defined(PSB_SERIAL_MPI)
-
  match_algorithm = ag%matching_alg
  n_sweeps        = ag%n_sweeps
  if (2**n_sweeps /= ag%orig_aggr_size) then
@@ -466,7 +460,7 @@ subroutine  amg_d_parmatch_aggregator_build_tprol(ag,parms,ag_data,&
    call psb_errpush(psb_err_from_subroutine_,name,a_err='amg_bootCMatch_if')
    goto 9999
  end if
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -110,11 +110,7 @@ subroutine amg_d_parmatch_smth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
  use amg_base_prec_type
  use amg_d_inner_mod
  use amg_d_base_aggregator_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_d_parmatch_aggregator_mod
-#else
  use amg_d_parmatch_aggregator_mod, amg_protect_name => amg_d_parmatch_smth_bld
-#endif
  implicit none

  ! Arguments
@@ -194,8 +190,6 @@ subroutine amg_d_parmatch_smth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&

  if (do_timings) call psb_tic(idx_phase1)

-#if !defined(PSB_SERIAL_MPI)
-
  naggr  = nlaggr(me+1)
  ntaggr = sum(nlaggr)
  naggrm1 = sum(nlaggr(1:me))
@@ -403,7 +397,7 @@ subroutine amg_d_parmatch_smth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
  if (debug_level >= psb_debug_outer_) &
       & write(debug_unit,*) me,' ',trim(name),&
       & 'Done smooth_aggregate '
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -99,11 +99,7 @@ subroutine amg_d_parmatch_spmm_bld(a,desc_a,ilaggr,nlaggr,parms,&
     & ac,desc_ac,op_prol,op_restr,t_prol,info)
  use psb_base_mod
  use amg_d_inner_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_d_parmatch_aggregator_mod
-#else
  use amg_d_parmatch_aggregator_mod, amg_protect_name => amg_d_parmatch_spmm_bld
-#endif
  implicit none

  ! Arguments
@@ -140,7 +136,6 @@ subroutine amg_d_parmatch_spmm_bld(a,desc_a,ilaggr,nlaggr,parms,&
  debug_unit  = psb_get_debug_unit()
  debug_level = psb_get_debug_level()

-#if !defined(PSB_SERIAL_MPI)
  call a%cp_to(acsr)

  call  amg_d_parmatch_spmm_bld_inner(acsr,desc_a,ilaggr,nlaggr,parms,&
@@ -154,7 +149,7 @@ subroutine amg_d_parmatch_spmm_bld(a,desc_a,ilaggr,nlaggr,parms,&
  if (debug_level >= psb_debug_outer_) &
       & write(debug_unit,*) me,' ',trim(name),&
       & 'Done spmm_bld '
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -99,11 +99,7 @@ subroutine amg_d_parmatch_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,&
     & ac,desc_ac,op_prol,op_restr,t_prol,info)
  use psb_base_mod
  use amg_d_inner_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_d_parmatch_aggregator_mod
-#else
  use amg_d_parmatch_aggregator_mod, amg_protect_name => amg_d_parmatch_spmm_bld_inner
-#endif
  implicit none

  ! Arguments
@@ -139,7 +135,6 @@ subroutine amg_d_parmatch_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,&
  info=psb_success_
  call psb_erractionsave(err_act)

-
  ictxt = desc_a%get_context()
  call psb_info(ictxt, me, np)
  debug_unit  = psb_get_debug_unit()
@@ -163,7 +158,6 @@ subroutine amg_d_parmatch_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,&
  naggrm1 = sum(nlaggr(1:me))
  naggrp1 = sum(nlaggr(1:me+1))

-#if !defined(PSB_SERIAL_MPI)
  !
  ! Here T_PROL should be arriving with GLOBAL indices on the cols
  ! and LOCAL indices on the rows.
@@ -203,11 +197,10 @@ subroutine amg_d_parmatch_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,&
    goto 9999
  end if

-
  if (debug_level >= psb_debug_outer_) &
       & write(debug_unit,*) me,' ',trim(name),&
       & 'Done smooth_aggregate '
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -99,11 +99,7 @@ subroutine amg_d_parmatch_spmm_bld_ov(a,desc_a,ilaggr,nlaggr,parms,&
     & ac,desc_ac,op_prol,op_restr,t_prol,info)
  use psb_base_mod
  use amg_d_inner_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_d_parmatch_aggregator_mod
-#else
  use amg_d_parmatch_aggregator_mod, amg_protect_name => amg_d_parmatch_spmm_bld_ov
-#endif
  implicit none

  ! Arguments
@@ -141,8 +137,6 @@ subroutine amg_d_parmatch_spmm_bld_ov(a,desc_a,ilaggr,nlaggr,parms,&
  debug_unit  = psb_get_debug_unit()
  debug_level = psb_get_debug_level()

-#if !defined(PSB_SERIAL_MPI)
-
  call a%mv_to(acsr)

  call  amg_d_parmatch_spmm_bld_inner(acsr,desc_a,ilaggr,nlaggr,parms,&
@@ -158,7 +152,6 @@ subroutine amg_d_parmatch_spmm_bld_ov(a,desc_a,ilaggr,nlaggr,parms,&
  if (debug_level >= psb_debug_outer_) &
       & write(debug_unit,*) me,' ',trim(name),&
       & 'Done spmm_bld '
-#endif
  call psb_erractionrestore(err_act)
  return

@@ -109,11 +109,7 @@ subroutine amg_d_parmatch_unsmth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
  use amg_base_prec_type
  use amg_d_inner_mod
  use amg_d_base_aggregator_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_d_parmatch_aggregator_mod
-#else
  use amg_d_parmatch_aggregator_mod, amg_protect_name => amg_d_parmatch_unsmth_bld
-#endif
  implicit none

  ! Arguments
@@ -168,7 +164,6 @@ subroutine amg_d_parmatch_unsmth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
    goto 9999
  end if

-#if !defined(PSB_SERIAL_MPI)
  nglob = desc_a%get_global_rows()
  nrow  = desc_a%get_local_rows()
  ncol  = desc_a%get_local_cols()
@@ -209,7 +204,8 @@ subroutine amg_d_parmatch_unsmth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
  call op_prol%mv_from(coo_prol)

  if (debug) write(0,*)  me,' ',trim(name),' After  mv_from',psb_get_errstatus()
-  if (debug) write(0,*)  me,' ',trim(name),' ',ac%get_fmt(),ac%get_nrows(),ac%get_ncols(),ac%get_nzeros(),naggr,ntaggr
+  if (debug) write(0,*)  me,' ',trim(name),' ',ac%get_fmt(),ac%get_nrows(),&
+       & ac%get_ncols(),ac%get_nzeros(),naggr,ntaggr
  ! write(0,*)  me,' ',trim(name),' Final AC newstyle ',ac%get_fmt(),ac%get_nrows(),ac%get_ncols(),ac%get_nzeros()

  if (debug) then
@@ -236,7 +232,6 @@ subroutine amg_d_parmatch_unsmth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
    goto 9999
  end if

-#endif
  call psb_erractionrestore(err_act)
  return

@@ -98,11 +98,7 @@ subroutine  amg_s_parmatch_aggregator_inner_mat_asb(ag,parms,a,desc_a,&
     & ac,desc_ac, op_prol,op_restr,info)
  use psb_base_mod
  use amg_base_prec_type
-#if defined(PSB_SERIAL_MPI)
-    use amg_s_parmatch_aggregator_mod
-#else
  use amg_s_parmatch_aggregator_mod, amg_protect_name => amg_s_parmatch_aggregator_inner_mat_asb
-#endif
  implicit none
  class(amg_s_parmatch_aggregator_type), target, intent(inout) :: ag
  type(amg_sml_parms), intent(inout)    :: parms
@@ -133,8 +129,6 @@ subroutine  amg_s_parmatch_aggregator_inner_mat_asb(ag,parms,a,desc_a,&
  ictxt = desc_a%get_context()
  call psb_info(ictxt,me,np)

-#if !defined(PSB_SERIAL_MPI)
-
  if (debug) write(0,*) me,' ',trim(name),' Start:',&
       & allocated(ag%ac),allocated(ag%desc_ac), allocated(ag%prol),allocated(ag%restr)

@@ -146,16 +140,17 @@ subroutine  amg_s_parmatch_aggregator_inner_mat_asb(ag,parms,a,desc_a,&
  case(amg_repl_mat_)
    !
    !
-    info = psb_err_internal_error_
-    call psb_errpush(info,name,a_err='no repl coarse_mat_ here')
-    goto 9999
-
+    if (np>1) then 
+      info = psb_err_internal_error_
+      call psb_errpush(info,name,a_err='no repl coarse_mat_ here')
+      goto 9999
+    end if
  case default
    info = psb_err_internal_error_
    call psb_errpush(info,name,a_err='invalid amg_coarse_mat_')
    goto 9999
  end select
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -98,11 +98,7 @@ subroutine  amg_s_parmatch_aggregator_mat_asb(ag,parms,a,desc_a,&
     & ac,desc_ac, op_prol,op_restr,info)
  use psb_base_mod
  use amg_base_prec_type
-#if defined(PSB_SERIAL_MPI)
-    use amg_s_parmatch_aggregator_mod
-#else
  use amg_s_parmatch_aggregator_mod, amg_protect_name => amg_s_parmatch_aggregator_mat_asb
-#endif
  implicit none
  class(amg_s_parmatch_aggregator_type), target, intent(inout) :: ag
  type(amg_sml_parms), intent(inout)    :: parms
@@ -135,8 +131,6 @@ subroutine  amg_s_parmatch_aggregator_mat_asb(ag,parms,a,desc_a,&
    return
  end if

-
-#if !defined(PSB_SERIAL_MPI)
  if (debug) write(0,*) me,' ',trim(name),' Start:',&
       & allocated(ag%ac),allocated(ag%desc_ac), allocated(ag%prol),allocated(ag%restr)

@@ -199,7 +193,7 @@ subroutine  amg_s_parmatch_aggregator_mat_asb(ag,parms,a,desc_a,&
    call psb_errpush(info,name,a_err='invalid amg_coarse_mat_')
    goto 9999
  end select
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -135,11 +135,7 @@ subroutine  amg_s_parmatch_aggregator_mat_bld(ag,parms,a,desc_a,ilaggr,nlaggr,&
  use psb_base_mod
  use amg_s_inner_mod
  use amg_base_prec_type
-#if defined(PSB_SERIAL_MPI)
-    use amg_s_parmatch_aggregator_mod
-#else
  use amg_s_parmatch_aggregator_mod, amg_protect_name => amg_s_parmatch_aggregator_mat_bld
-#endif
  implicit none

  class(amg_s_parmatch_aggregator_type), target, intent(inout) :: ag
@@ -176,7 +172,6 @@ subroutine  amg_s_parmatch_aggregator_mat_bld(ag,parms,a,desc_a,ilaggr,nlaggr,&
  ! algorithm specified by
  !

-#if !defined(PSB_SERIAL_MPI)
  call clean_shortcuts(ag)
  !
  ! When requesting smoothed aggregation we cannot use the
@@ -212,14 +207,12 @@ subroutine  amg_s_parmatch_aggregator_mat_bld(ag,parms,a,desc_a,ilaggr,nlaggr,&
    call psb_errpush(psb_err_from_subroutine_,name,a_err='Inner aggrmat asb')
    goto 9999
  end if
-#endif
  call psb_erractionrestore(err_act)
  return

 9999 call psb_error_handler(err_act)
  return

-#if !defined(PSB_SERIAL_MPI)

 contains
  subroutine clean_shortcuts(ag)
@@ -248,5 +241,4 @@ contains
      end if
    end if
  end subroutine clean_shortcuts
-#endif
 end subroutine amg_s_parmatch_aggregator_mat_bld
@@ -47,11 +47,7 @@ subroutine  amg_s_parmatch_aggregator_build_tprol(ag,parms,ag_data,&
  use psb_base_mod
  use amg_base_prec_type
  use amg_s_inner_mod
-#if defined(PSB_SERIAL_MPI)
-  use amg_s_parmatch_aggregator_mod
-#else
  use amg_s_parmatch_aggregator_mod, amg_protect_name => amg_s_parmatch_aggregator_build_tprol
-#endif
  use iso_c_binding
  implicit none
  class(amg_s_parmatch_aggregator_type), target, intent(inout) :: ag
@@ -120,8 +116,6 @@ subroutine  amg_s_parmatch_aggregator_build_tprol(ag,parms,ag_data,&
       &   amg_aggr_ord_nat_,is_legal_ml_aggr_ord)
  call amg_check_def(parms%aggr_thresh,'Aggr_Thresh',szero,is_legal_s_aggr_thrs)

-#if !defined(PSB_SERIAL_MPI)
-
  match_algorithm = ag%matching_alg
  n_sweeps        = ag%n_sweeps
  if (2**n_sweeps /= ag%orig_aggr_size) then
@@ -466,7 +460,7 @@ subroutine  amg_s_parmatch_aggregator_build_tprol(ag,parms,ag_data,&
    call psb_errpush(psb_err_from_subroutine_,name,a_err='amg_bootCMatch_if')
    goto 9999
  end if
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -110,11 +110,7 @@ subroutine amg_s_parmatch_smth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
  use amg_base_prec_type
  use amg_s_inner_mod
  use amg_s_base_aggregator_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_s_parmatch_aggregator_mod
-#else
  use amg_s_parmatch_aggregator_mod, amg_protect_name => amg_s_parmatch_smth_bld
-#endif
  implicit none

  ! Arguments
@@ -194,8 +190,6 @@ subroutine amg_s_parmatch_smth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&

  if (do_timings) call psb_tic(idx_phase1)

-#if !defined(PSB_SERIAL_MPI)
-
  naggr  = nlaggr(me+1)
  ntaggr = sum(nlaggr)
  naggrm1 = sum(nlaggr(1:me))
@@ -403,7 +397,7 @@ subroutine amg_s_parmatch_smth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
  if (debug_level >= psb_debug_outer_) &
       & write(debug_unit,*) me,' ',trim(name),&
       & 'Done smooth_aggregate '
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -99,11 +99,7 @@ subroutine amg_s_parmatch_spmm_bld(a,desc_a,ilaggr,nlaggr,parms,&
     & ac,desc_ac,op_prol,op_restr,t_prol,info)
  use psb_base_mod
  use amg_s_inner_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_s_parmatch_aggregator_mod
-#else
  use amg_s_parmatch_aggregator_mod, amg_protect_name => amg_s_parmatch_spmm_bld
-#endif
  implicit none

  ! Arguments
@@ -140,7 +136,6 @@ subroutine amg_s_parmatch_spmm_bld(a,desc_a,ilaggr,nlaggr,parms,&
  debug_unit  = psb_get_debug_unit()
  debug_level = psb_get_debug_level()

-#if !defined(PSB_SERIAL_MPI)
  call a%cp_to(acsr)

  call  amg_s_parmatch_spmm_bld_inner(acsr,desc_a,ilaggr,nlaggr,parms,&
@@ -154,7 +149,7 @@ subroutine amg_s_parmatch_spmm_bld(a,desc_a,ilaggr,nlaggr,parms,&
  if (debug_level >= psb_debug_outer_) &
       & write(debug_unit,*) me,' ',trim(name),&
       & 'Done spmm_bld '
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -99,11 +99,7 @@ subroutine amg_s_parmatch_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,&
     & ac,desc_ac,op_prol,op_restr,t_prol,info)
  use psb_base_mod
  use amg_s_inner_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_s_parmatch_aggregator_mod
-#else
  use amg_s_parmatch_aggregator_mod, amg_protect_name => amg_s_parmatch_spmm_bld_inner
-#endif
  implicit none

  ! Arguments
@@ -139,7 +135,6 @@ subroutine amg_s_parmatch_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,&
  info=psb_success_
  call psb_erractionsave(err_act)

-
  ictxt = desc_a%get_context()
  call psb_info(ictxt, me, np)
  debug_unit  = psb_get_debug_unit()
@@ -163,7 +158,6 @@ subroutine amg_s_parmatch_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,&
  naggrm1 = sum(nlaggr(1:me))
  naggrp1 = sum(nlaggr(1:me+1))

-#if !defined(PSB_SERIAL_MPI)
  !
  ! Here T_PROL should be arriving with GLOBAL indices on the cols
  ! and LOCAL indices on the rows.
@@ -203,11 +197,10 @@ subroutine amg_s_parmatch_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,&
    goto 9999
  end if

-
  if (debug_level >= psb_debug_outer_) &
       & write(debug_unit,*) me,' ',trim(name),&
       & 'Done smooth_aggregate '
-#endif
+
  call psb_erractionrestore(err_act)
  return

@@ -99,11 +99,7 @@ subroutine amg_s_parmatch_spmm_bld_ov(a,desc_a,ilaggr,nlaggr,parms,&
     & ac,desc_ac,op_prol,op_restr,t_prol,info)
  use psb_base_mod
  use amg_s_inner_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_s_parmatch_aggregator_mod
-#else
  use amg_s_parmatch_aggregator_mod, amg_protect_name => amg_s_parmatch_spmm_bld_ov
-#endif
  implicit none

  ! Arguments
@@ -141,8 +137,6 @@ subroutine amg_s_parmatch_spmm_bld_ov(a,desc_a,ilaggr,nlaggr,parms,&
  debug_unit  = psb_get_debug_unit()
  debug_level = psb_get_debug_level()

-#if !defined(PSB_SERIAL_MPI)
-
  call a%mv_to(acsr)

  call  amg_s_parmatch_spmm_bld_inner(acsr,desc_a,ilaggr,nlaggr,parms,&
@@ -158,7 +152,6 @@ subroutine amg_s_parmatch_spmm_bld_ov(a,desc_a,ilaggr,nlaggr,parms,&
  if (debug_level >= psb_debug_outer_) &
       & write(debug_unit,*) me,' ',trim(name),&
       & 'Done spmm_bld '
-#endif
  call psb_erractionrestore(err_act)
  return

@@ -109,11 +109,7 @@ subroutine amg_s_parmatch_unsmth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
  use amg_base_prec_type
  use amg_s_inner_mod
  use amg_s_base_aggregator_mod
-#if defined(PSB_SERIAL_MPI)
-    use amg_s_parmatch_aggregator_mod
-#else
  use amg_s_parmatch_aggregator_mod, amg_protect_name => amg_s_parmatch_unsmth_bld
-#endif
  implicit none

  ! Arguments
@@ -168,7 +164,6 @@ subroutine amg_s_parmatch_unsmth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
    goto 9999
  end if

-#if !defined(PSB_SERIAL_MPI)
  nglob = desc_a%get_global_rows()
  nrow  = desc_a%get_local_rows()
  ncol  = desc_a%get_local_cols()
@@ -209,7 +204,8 @@ subroutine amg_s_parmatch_unsmth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
  call op_prol%mv_from(coo_prol)

  if (debug) write(0,*)  me,' ',trim(name),' After  mv_from',psb_get_errstatus()
-  if (debug) write(0,*)  me,' ',trim(name),' ',ac%get_fmt(),ac%get_nrows(),ac%get_ncols(),ac%get_nzeros(),naggr,ntaggr
+  if (debug) write(0,*)  me,' ',trim(name),' ',ac%get_fmt(),ac%get_nrows(),&
+       & ac%get_ncols(),ac%get_nzeros(),naggr,ntaggr
  ! write(0,*)  me,' ',trim(name),' Final AC newstyle ',ac%get_fmt(),ac%get_nrows(),ac%get_ncols(),ac%get_nzeros()

  if (debug) then
@@ -236,7 +232,6 @@ subroutine amg_s_parmatch_unsmth_bld(dol1smoothing,ag,a,desc_a,ilaggr,nlaggr,&
    goto 9999
  end if

-#endif
  call psb_erractionrestore(err_act)
  return

@@ -1,6 +1,6 @@
 #include "MatchBoxPC.h"
 // TODO comment
-#if !defined(SERIAL_MPI)
+#if !defined(PSB_SERIAL_MPI)

 void clean(MilanLongInt NLVer,
           MilanInt myRank,
@@ -8,6 +8,7 @@
 * @param edgeLocWeight
 * @return
 */
+#if !defined(PSB_SERIAL_MPI)

 MilanLongInt firstComputeCandidateMateD(MilanLongInt adj1,
                                       MilanLongInt adj2,
@@ -136,3 +137,4 @@ MilanLongInt computeCandidateMateS(MilanLongInt adj1,
    return w;
 }

+#endif
@@ -41,7 +41,10 @@

 #ifndef _static_Queue_
 #define _static_Queue_
+#include "amg_config.h"

+
+#if !defined(PSB_SERIAL_MPI)
 #include "primitiveDataTypeDefinitions.h"
 #include "preProcessorDirectives.h"

@@ -199,3 +202,4 @@ class staticQueue
 };

 #endif
+#endif
@@ -1,4 +1,8 @@
+#include "amg_config.h"
 #include "MatchBoxPC.h"
+
+#if !defined(PSB_SERIAL_MPI)
+
 void extractUChunk(
    vector<MilanLongInt> &UChunkBeingProcessed,
    vector<MilanLongInt> &U,
@@ -28,3 +32,4 @@ void extractUChunk(

    } // End of critical U // End of critical U
 }
+#endif
@@ -1,4 +1,6 @@
+#include "amg_config.h"
 #include "MatchBoxPC.h"
+#if !defined(PSB_SERIAL_MPI)
 /// Find the owner of a ghost node:
 MilanInt findOwnerOfGhost(MilanLongInt vtxIndex, MilanLongInt *mVerDistance,
                          MilanInt myRank, MilanInt numProcs)
@@ -26,3 +28,4 @@ MilanInt findOwnerOfGhost(MilanLongInt vtxIndex, MilanLongInt *mVerDistance,

  return Current;
 } // End of findOwnerOfGhost()
+#endif
@@ -1,4 +1,7 @@
+#include "amg_config.h"
 #include "MatchBoxPC.h"
+#if !defined(PSB_SERIAL_MPI)
+
 void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
                MilanLongInt StartIndex, MilanLongInt EndIndex,
                MilanLongInt *numGhostEdges,
@@ -301,3 +304,4 @@ void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
        } // End of single region
    }     // End of parallel region
 }
+#endif
@@ -1,4 +1,7 @@
+#include "amg_config.h"
 #include "MatchBoxPC.h"
+#if !defined(PSB_SERIAL_MPI)
+
 /**
 * //TODO documentation
 * @param k
@@ -43,3 +46,4 @@ bool isAlreadyMatched(MilanLongInt node,

    return val >= 0; // Already matched
 }
+#endif
@@ -1,4 +1,6 @@
+#include "amg_config.h"
 #include "MatchBoxPC.h"
+#if !defined(PSB_SERIAL_MPI)

 void PARALLEL_COMPUTE_CANDIDATE_MATE_BD(MilanLongInt NLVer,
                                              MilanLongInt *verLocPtr,
@@ -53,4 +55,4 @@ void PARALLEL_COMPUTE_CANDIDATE_MATE_BS(MilanLongInt NLVer,
        }
    }
 }
-
+#endif
@@ -41,6 +41,8 @@

 #ifndef _preprocessor_Directives_
 #define _preprocessor_Directives_
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)

 //I/O
 #include <iostream>
@@ -72,10 +74,10 @@
 #include <map>

 //MPI:
-#if !defined(SERIAL_MPI)
+#if !defined(PSB_SERIAL_MPI)
 #include "mpi.h"
 #endif


-
+#endif
 #endif
@@ -41,6 +41,8 @@

 #ifndef _primitiveDataType_Definition_
 #define _primitiveDataType_Definition_
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)

 #include "preProcessorDirectives.h"

@@ -152,5 +154,5 @@ const float FMINUS_INFINITY = -FPLUS_INFINITY;


 #endif
-
+#endif
 #endif
@@ -1,3 +1,5 @@
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)
 #include "MatchBoxPC.h"
 void PROCESS_CROSS_EDGE(MilanLongInt *edge,
                        MilanLongInt *S)
@@ -21,3 +23,4 @@ void PROCESS_CROSS_EDGE(MilanLongInt *edge,

    // End: PARALLEL_PROCESS_CROSS_EDGE_B
 }
+#endif
@@ -1,3 +1,5 @@
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)
 #include "MatchBoxPC.h" 
 void PARALLEL_PROCESS_EXPOSED_VERTEX_BD(MilanLongInt NLVer,
                                       MilanLongInt *candidateMate,
@@ -367,3 +369,4 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_BS(MilanLongInt NLVer,
    } // End of parallel region
 }

+#endif
@@ -1,3 +1,5 @@
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)
 #include "MatchBoxPC.h"
 void processMatchedVerticesD(
    MilanLongInt NLVer,
@@ -583,3 +585,4 @@ void processMatchedVerticesS(
    } // End of parallel region
 }

+#endif
@@ -1,3 +1,5 @@
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)
 #include "MatchBoxPC.h"
 //#define DEBUG_HANG_
 void processMatchedVerticesAndSendMessagesD(
@@ -589,3 +591,4 @@ void processMatchedVerticesAndSendMessagesS(
 #endif
 }

+#endif
@@ -1,6 +1,7 @@
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)
 #include "MatchBoxPC.h"
 //#define DEBUG_HANG_
-#if !defined(SERIAL_MPI)

 void processMessagesD(
    MilanLongInt NLVer,
@@ -1,3 +1,5 @@
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)
 #include "MatchBoxPC.h"
 void queuesTransfer(vector<MilanLongInt> &U,
                    vector<MilanLongInt> &privateU,
@@ -31,3 +33,4 @@ void queuesTransfer(vector<MilanLongInt> &U,

 }

+#endif
@@ -1,3 +1,5 @@
+#include "amg_config.h"
+#if !defined(PSB_SERIAL_MPI)
 #include "MatchBoxPC.h"
 void sendBundledMessages(MilanLongInt *numGhostEdges,
 			 MilanInt *BufferSize,
@@ -154,7 +156,7 @@ void sendBundledMessages(MilanLongInt *numGhostEdges,

 #pragma omp task depend(inout : OneMessageSize, *BufferSize) depend(out : numMessagesToSend) depend(in : *numGhostEdges)
 {
-
+  
 #ifdef PRINT_DEBUG_INFO_
  cout << "\n(" << myRank << ")Number of Ghost edges = " << *numGhostEdges;
  cout << "\n(" << myRank << ")Total number of potential message X 2 = " << *numGhostEdges * 2;
@@ -206,3 +208,4 @@ void sendBundledMessages(MilanLongInt *numGhostEdges,
 }
 }
 }
+#endif
@@ -252,15 +252,9 @@ subroutine amg_c_base_onelev_csetc(lv,what,val,info,pos,idx)
      allocate(amg_c_dec_aggregator_type :: lv%aggr, stat=info)
    case('SYMDEC')
      allocate(amg_c_symdec_aggregator_type :: lv%aggr, stat=info)
-#if !defined(PSB_SERIAL_MPI)
-#endif
  case default
      info =  psb_err_internal_error_
-#if !defined(PSB_SERIAL_MPI)
      call psb_errpush(info,name,a_err='Unsupported PAR_AGGR_ALG')
-#else
-      call psb_errpush(info,name,a_err='PAR_AGGR_ALG unsupported (PSB_SERIAL_MPI on)')
-#endif
      goto 9999
    end select
    if (info == psb_success_) call lv%aggr%default()
@@ -42,9 +42,7 @@ subroutine amg_d_base_onelev_csetc(lv,what,val,info,pos,idx)
  use amg_d_base_aggregator_mod
  use amg_d_dec_aggregator_mod
  use amg_d_symdec_aggregator_mod
-#if !defined(PSB_SERIAL_MPI)
  use amg_d_parmatch_aggregator_mod
-#endif
  use amg_d_poly_smoother
  use amg_d_jac_smoother
  use amg_d_as_smoother
@@ -280,17 +278,11 @@ subroutine amg_d_base_onelev_csetc(lv,what,val,info,pos,idx)
      allocate(amg_d_dec_aggregator_type :: lv%aggr, stat=info)
    case('SYMDEC')
      allocate(amg_d_symdec_aggregator_type :: lv%aggr, stat=info)
-#if !defined(PSB_SERIAL_MPI)
    case('COUP','COUPLED')
      allocate(amg_d_parmatch_aggregator_type :: lv%aggr, stat=info)
-#endif
  case default
      info =  psb_err_internal_error_
-#if !defined(PSB_SERIAL_MPI)
      call psb_errpush(info,name,a_err='Unsupported PAR_AGGR_ALG')
-#else
-      call psb_errpush(info,name,a_err='PAR_AGGR_ALG unsupported (PSB_SERIAL_MPI on)')
-#endif
      goto 9999
    end select
    if (info == psb_success_) call lv%aggr%default()
@@ -42,9 +42,7 @@ subroutine amg_s_base_onelev_csetc(lv,what,val,info,pos,idx)
  use amg_s_base_aggregator_mod
  use amg_s_dec_aggregator_mod
  use amg_s_symdec_aggregator_mod
-#if !defined(PSB_SERIAL_MPI)
  use amg_s_parmatch_aggregator_mod
-#endif
  use amg_s_poly_smoother
  use amg_s_jac_smoother
  use amg_s_as_smoother
@@ -260,17 +258,11 @@ subroutine amg_s_base_onelev_csetc(lv,what,val,info,pos,idx)
      allocate(amg_s_dec_aggregator_type :: lv%aggr, stat=info)
    case('SYMDEC')
      allocate(amg_s_symdec_aggregator_type :: lv%aggr, stat=info)
-#if !defined(PSB_SERIAL_MPI)
    case('COUP','COUPLED')
      allocate(amg_s_parmatch_aggregator_type :: lv%aggr, stat=info)
-#endif
  case default
      info =  psb_err_internal_error_
-#if !defined(PSB_SERIAL_MPI)
      call psb_errpush(info,name,a_err='Unsupported PAR_AGGR_ALG')
-#else
-      call psb_errpush(info,name,a_err='PAR_AGGR_ALG unsupported (PSB_SERIAL_MPI on)')
-#endif
      goto 9999
    end select
    if (info == psb_success_) call lv%aggr%default()
@@ -272,15 +272,9 @@ subroutine amg_z_base_onelev_csetc(lv,what,val,info,pos,idx)
      allocate(amg_z_dec_aggregator_type :: lv%aggr, stat=info)
    case('SYMDEC')
      allocate(amg_z_symdec_aggregator_type :: lv%aggr, stat=info)
-#if !defined(PSB_SERIAL_MPI)
-#endif
  case default
      info =  psb_err_internal_error_
-#if !defined(PSB_SERIAL_MPI)
      call psb_errpush(info,name,a_err='Unsupported PAR_AGGR_ALG')
-#else
-      call psb_errpush(info,name,a_err='PAR_AGGR_ALG unsupported (PSB_SERIAL_MPI on)')
-#endif
      goto 9999
    end select
    if (info == psb_success_) call lv%aggr%default()
@@ -58,8 +58,8 @@ VCYCLE                      ! Type of multilevel CYCLE: VCYCLE WCYCLE KCYCLE MUL
 -3                          ! Max Number of levels in a multilevel preconditioner; if <0, lib default
 -3                          ! Target coarse matrix size per process; if <0, lib default
 SMOOTHED                    ! Type of aggregation: SMOOTHED UNSMOOTHED
-DEC                         ! Parallel aggregation: DEC, SYMDEC, COUPLED
-SOC1                        ! aggregation measure SOC1, MATCHBOXP
+COUPLED                     ! Parallel aggregation: DEC, SYMDEC, COUPLED
+MATCHBOXP                   ! aggregation measure SOC1, MATCHBOXP
 8                           ! Requested size of the aggregates for MATCHBOXP
 NATURAL                     ! Ordering of aggregation NATURAL DEGREE
 -1.5                        ! Coarsening ratio, if < 0 use library default