fdurastante/amg4psblas
1
0
Fork 0
mirror of https://github.com/sfilippone/amg4psblas.git synced 2026-10-19 20:32:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

Modify MATCHBOXP to fix OpenMP. Performance to be reviewed

Browse Source
This commit is contained in:
Salvatore Filippone
2023-02-10 07:50:58 -05:00
parent 32994c7ce8
commit ebe9b45177
9 changed files with 979 additions and 962 deletions
Download Patch File Download Diff File Expand all files Collapse all files
amgprec/impl/aggregator/processExposedVertex.cpp
+72 -82
View File
@@ -29,23 +29,23 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner)
{
{
MilanLongInt v = -1, k = -1, w = -1, adj11 = 0, adj12 = 0, k1 = 0;
MilanInt ghostOwner = 0, option;
MilanInt ghostOwner = 0, option, igw;
#pragma omp parallel private(option, k, w, v, k1, adj11, adj12, ghostOwner) \
firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx, privateQMsgType, privateQOwner) default(shared) \
num_threads(NUM_THREAD)
#pragma omp parallel private(option, k, w, v, k1, adj11, adj12, ghostOwner) \
firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx, privateQMsgType, privateQOwner) \
default(shared) num_threads(NUM_THREAD)
{
#pragma omp for reduction(+ \
: PCounter[:numProcs], myCard \
[:1], msgInd \
[:1], NumMessagesBundled \
[:1]) schedule(static)
for (v = 0; v < NLVer; v++)
{
[:1]) \
schedule(static)
for (v = 0; v < NLVer; v++) {
option = -1;
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
k = candidateMate[v];
@@ -67,91 +67,81 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
#pragma omp critical(processExposed)
{
if (isAlreadyMatched(verLocInd[k], StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap))
{
w = computeCandidateMate(verLocPtr[v],
verLocPtr[v + 1],
edgeLocWeight, 0,
verLocInd,
StartIndex,
EndIndex,
GMate,
Mate,
Ghost2LocalMap);
candidateMate[v] = w;
if (isAlreadyMatched(verLocInd[k], StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap)) {
w = computeCandidateMate(verLocPtr[v],
verLocPtr[v + 1],
edgeLocWeight, 0,
verLocInd,
StartIndex,
EndIndex,
GMate,
Mate,
Ghost2LocalMap);
candidateMate[v] = w;
}
if (w >= 0) {
(*myCard)++;
if ((w < StartIndex) || (w > EndIndex)) { // w is a ghost vertex
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v + StartIndex) {
option = 1;
Mate[v] = w;
GMate[Ghost2LocalMap[w]] = v + StartIndex; // w is a Ghost
} // End of if CandidateMate[w] = v
if (w >= 0)
{
(*myCard)++;
if ((w < StartIndex) || (w > EndIndex))
{ // w is a ghost vertex
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v + StartIndex)
{
option = 1;
Mate[v] = w;
GMate[Ghost2LocalMap[w]] = v + StartIndex; // w is a Ghost
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else
{ // w is a local vertex
if (candidateMate[w - StartIndex] == (v + StartIndex))
{
option = 3;
Mate[v] = w; // v is local
Mate[w - StartIndex] = v + StartIndex; // w is local
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == (v + StartIndex)) {
option = 3;
Mate[v] = w; // v is local
Mate[w - StartIndex] = v + StartIndex; // w is local
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v + StartIndex << "," << w << ") ";
fflush(stdout);
cout << "\n(" << myRank << ")MATCH: (" << v + StartIndex << "," << w << ") ";
fflush(stdout);
#endif
} // End of if ( candidateMate[w-StartIndex] == (v+StartIndex) )
} // End of Else
} // End of if ( candidateMate[w-StartIndex] == (v+StartIndex) )
} // End of Else
} // End of second if
} // End critical processExposed
} // End of if(w >=0)
else
{
// This piece of code is executed a really small amount of times
adj11 = verLocPtr[v];
adj12 = verLocPtr[v + 1];
for (k1 = adj11; k1 < adj12; k1++)
{
w = verLocInd[k1];
if ((w < StartIndex) || (w > EndIndex))
{ // A ghost
else {
// This piece of code is executed a really small amount of times
adj11 = verLocPtr[v];
adj12 = verLocPtr[v + 1];
for (k1 = adj11; k1 < adj12; k1++) {
w = verLocInd[k1];
if ((w < StartIndex) || (w > EndIndex)) { // A ghost
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a failure message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
fflush(stdout);
cout << "\n(" << myRank << ")Sending a failure message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
fflush(stdout);
#endif
(*msgInd)++;
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
PCounter[ghostOwner]++;
privateQLocalVtx.push_back(v + StartIndex);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
(*msgInd)++;
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
PCounter[ghostOwner]++;
privateQLocalVtx.push_back(v + StartIndex);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
}
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
switch (option)
{
case -1:
@@ -202,4 +192,4 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
privateQOwner);
} // End of parallel region
}
}