fdurastante
/
amg4psblas
mirror of https://github.com/sfilippone/amg4psblas.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Single parallel regions with multiple for cycles

Browse Source 
Added OMP for testing
omp-walther
StefanoPetrilli 3 years ago
parent 8f6dc5fac2
commit 0a8debe43a
4 changed files with 115 additions and 83 deletions
Show Stats Download Patch File Download Diff File

2
amgprec/impl/aggregator/Makefile
Unescape Escape View File

@ -4,7 +4,7 @@ INCDIR=../../../include
MODDIR=../../../modules
HERE=../..
FINCLUDES=$(FMFLAG)$(HERE) $(FMFLAG)$(MODDIR) $(FMFLAG)$(INCDIR) $(PSBLAS_INCLUDES) -fopenmp
FINCLUDES=$(FMFLAG)$(HERE) $(FMFLAG)$(MODDIR) $(FMFLAG)$(INCDIR) $(PSBLAS_INCLUDES)
CXXINCLUDES=$(FMFLAG)$(HERE) $(FMFLAG)$(INCDIR) $(FMFLAG)/.
#CINCLUDES= -I${SUPERLU_INCDIR} -I${HSL_INCDIR} -I${SPRAL_INCDIR} -I/home/users/pasqua/Ambra/BootCMatch/include -lBCM -L/home/users/pasqua/Ambra/BootCMatch/lib -lm

3
amgprec/impl/aggregator/MatchBoxPC.cpp
Unescape Escape View File

@ -66,6 +66,7 @@ void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
myRank,NLVer, NLEdge,verDistance[0],verDistance[1]);
#endif
#define TIME_TRACKER
#ifdef TIME_TRACKER
double tmr = MPI_Wtime();
#endif
@ -80,7 +81,7 @@ void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
#ifdef TIME_TRACKER
tmr = MPI_Wtime() - tmr;
fprintf(stderr, "Elaboration time: %f\n", tmr);
fprintf(stderr, "Elaboration time: %f for $ld\n", tmr, NLEdge);
#endif
#endif

55
amgprec/impl/aggregator/algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC.cpp
Unescape Escape View File

@ -180,17 +180,34 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
* only when a ghost edge is found and ghost edges are a minority.
*/
//Define Adjacency Lists for Ghost Vertices:
//cout<<"Building Ghost data structures ... \n\n";
vector <MilanLongInt> verGhostPtr, verGhostInd, tempCounter;
//Mate array for ghost vertices:
vector <MilanLongInt> GMate; //Proportional to the number of ghost vertices
#ifdef TIME_TRACKER
double Ghost2LocalInitialization = MPI_Wtime();
#endif
#pragma omp parallel for private(insertMe) firstprivate(StartIndex, EndIndex) default(shared)
//#define OMP
#ifdef OMP
#pragma omp parallel private(insertMe, k, adj1, adj2) firstprivate(StartIndex, EndIndex) default(shared) num_threads(4)
{
#endif
//printf("Id %d\n", omp_get_thread_num());
#ifdef OMP
#pragma omp for
#endif
for (i = 0; i < NLEdge; i++) { //O(m) - Each edge stored twice
insertMe = verLocInd[i];
//cout<<"InsertMe on Process "<<myRank<<" is: "<<insertMe<<endl;
if ((insertMe < StartIndex) || (insertMe > EndIndex)) { //Find a ghost
#ifdef OMP
#pragma omp critical
{
#endif
numGhostEdges++;
storedAlready = Ghost2LocalMap.find(insertMe);
if (storedAlready != Ghost2LocalMap.end()) { //Has already been added
@ -202,10 +219,17 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
Counter.push_back(1); //Initialize the counter
numGhostVertices++; //Increment the number of ghost vertices
} //End of else()
#ifdef OMP
}
#endif
} //End of if ( (insertMe < StartIndex) || (insertMe > EndIndex) )
} //End of for(ghost vertices)
#ifdef OMP
#pragma omp single
{
#endif
#ifdef TIME_TRACKER
Ghost2LocalInitialization = MPI_Wtime() - Ghost2LocalInitialization;
fprintf(stderr, "Ghost2LocalInitialization time: %f\n", Ghost2LocalInitialization);
@ -224,11 +248,8 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
} while ( storedAlready != Ghost2LocalMap.end() );
}
#endif
//Build Adjacency Lists for Ghost Vertices:
//cout<<"Building Ghost data structures ... \n\n";
vector<MilanLongInt> verGhostPtr, verGhostInd, tempCounter;
//Mate array for ghost vertices:
vector<MilanLongInt> GMate; //Proportional to the number of ghost vertices
//Initialize adjacency Lists for Ghost Vertices:
try {
verGhostPtr.reserve(numGhostVertices + 1); //Pointer Vector
tempCounter.reserve(numGhostVertices); //Pointer Vector
@ -249,17 +270,21 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
cout<<"\n("<<myRank<<")Ghost Vertex Pointer: "; fflush(stdout);
#endif
#define TIME_TRACKER
#ifdef TIME_TRACKER
double verGhostPtrInitialization = MPI_Wtime();
#endif
#ifdef OMP
}
#endif
/*
* OMP verGhostPtrInitialization
*
*/
#pragma omp parallel for default(shared)
#ifdef OMP
#pragma omp for nowait
#endif
for (i = 0; i < numGhostVertices; i++) { //O(|Ghost Vertices|)
verGhostPtr[i + 1] = verGhostPtr[i] + Counter[i];
#ifdef PRINT_DEBUG_INFO_
@ -272,9 +297,6 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
fprintf(stderr, "verGhostPtrInitialization time: %f\n", verGhostPtrInitialization);
#endif
#undef TIME_TRACKER
#ifdef PRINT_DEBUG_INFO_
if ( numGhostVertices > 0 )
cout<<verGhostPtr[numGhostVertices]<<"\n";
@ -299,24 +321,33 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
#ifdef TIME_TRACKER
double verGhostIndInitialization = MPI_Wtime();
#endif
#pragma omp parallel for private(insertMe, k, adj1, adj2) firstprivate(StartIndex, EndIndex) default(shared)
#ifdef OMP
#pragma omp for
#endif
for ( v=0; v < NLVer; v++ ) {
adj1 = verLocPtr[v]; //Vertex Pointer
adj2 = verLocPtr[v+1];
for( k = adj1; k < adj2; k++ ) {
w = verLocInd[k]; //Get the adjacent vertex
if ( (w < StartIndex) || (w > EndIndex) ) { //Find a ghost
#ifdef OMP
#pragma omp critical
{
#endif
insertMe = verGhostPtr[Ghost2LocalMap[w]] + tempCounter[Ghost2LocalMap[w]]; //Where to insert
verGhostInd[insertMe] = v + StartIndex; //Add the adjacency
tempCounter[Ghost2LocalMap[w]]++; //Increment the counter
#ifdef OMP
}
#endif
} //End of if((w < StartIndex) || (w > EndIndex))
} //End of for(k)
} //End of for (v)
tempCounter.clear(); //Do not need this any more
#ifdef OMP
} //end of parallel region
#endif
#ifdef TIME_TRACKER
verGhostIndInitialization = MPI_Wtime() - verGhostIndInitialization;
fprintf(stderr, "verGhostIndInitialization time: %f\n", verGhostIndInitialization);

2
exec.sh
Unescape Escape View File

@ -2,6 +2,6 @@ make all
cd samples/advanced/pdegen
make amg_d_pde3d
cd runs
mpirun -np 8 amg_d_pde3d amg_pde3d.inp
mpirun -np 2 amg_d_pde3d amg_pde3d.inp

Write Preview
Loading…
Cancel
Save