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

processMatchedVertices parallelized

Browse Source
omp-walther
StefanoPetrilli 2 years ago
parent 71d4cdc319
commit 9ab54adf3f
1 changed files with 166 additions and 148 deletions
Show Stats Download Patch File Download Diff File

308
amgprec/impl/aggregator/processMatchedVertices.cpp
Unescape Escape View File

@ -36,6 +36,7 @@ void processMatchedVertices(
{ {
MilanLongInt adj1, adj2, adj11, adj12, k, k1, v = -1, w = -1, ghostOwner; MilanLongInt adj1, adj2, adj11, adj12, k, k1, v = -1, w = -1, ghostOwner;
int option;
#ifdef PRINT_DEBUG_INFO_ #ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl; cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout); fflush(stdout);
@ -45,56 +46,55 @@ void processMatchedVertices(
#ifdef COUNT_LOCAL_VERTEX #ifdef COUNT_LOCAL_VERTEX
MilanLongInt localVertices = 0; MilanLongInt localVertices = 0;
#endif #endif
#pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner) firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx, privateQMsgType, privateQOwner) default(shared) num_threads(NUM_THREAD) #pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner, option) firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx, privateQMsgType, privateQOwner) default(shared) num_threads(NUM_THREAD)
{ {
// TODO what would be the optimal UCHUNK // TODO what would be the optimal UCHUNK
// TODO refactor // TODO refactor
vector<MilanLongInt> UChunkBeingProcessed; vector<MilanLongInt> UChunkBeingProcessed;
UChunkBeingProcessed.reserve(UCHUNK); UChunkBeingProcessed.reserve(UCHUNK);
while (!U.empty()) while (!U.empty())
{ {
extractUChunk(UChunkBeingProcessed, U, privateU); extractUChunk(UChunkBeingProcessed, U, privateU);
for (MilanLongInt u : UChunkBeingProcessed) for (MilanLongInt u : UChunkBeingProcessed)
{ {
#ifdef PRINT_DEBUG_INFO_ #ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")u: " << u; cout << "\n(" << myRank << ")u: " << u;
fflush(stdout); fflush(stdout);
#endif #endif
if ((u >= StartIndex) && (u <= EndIndex)) if ((u >= StartIndex) && (u <= EndIndex))
{ // Process Only the Local Vertices { // Process Only the Local Vertices
#ifdef COUNT_LOCAL_VERTEX #ifdef COUNT_LOCAL_VERTEX
localVertices++; localVertices++;
#endif #endif
// Get the Adjacency list for u // Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1]; adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++) for (k = adj1; k < adj2; k++)
{ {
v = verLocInd[k]; option = -1;
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex)) if ((v >= StartIndex) && (v <= EndIndex))
{ // If Local Vertex: { // If Local Vertex:
#ifdef PRINT_DEBUG_INFO_ #ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v]; cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout); fflush(stdout);
#endif #endif
// If the current vertex is pointing to a matched vertex and is not matched // If the current vertex is pointing to a matched vertex and is not matched
// FIXME is there a way to make candidateMate private? if (not isAlreadyMatched(v, StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap))
// for the moment it could generate an error. {
if (not isAlreadyMatched(v, StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap))
{
#pragma omp critical #pragma omp critical
{
if (candidateMate[v - StartIndex] == u)
{ {
if (candidateMate[v - StartIndex] == u)
{
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v) // Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMate(verLocPtr[v - StartIndex], w = computeCandidateMate(verLocPtr[v - StartIndex],
verLocPtr[v - StartIndex + 1], verLocPtr[v - StartIndex + 1],
@ -122,171 +122,189 @@ void processMatchedVertices(
cout << "\n(" << myRank << ")Sending a request message:"; cout << "\n(" << myRank << ")Sending a request message:";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs); cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
#endif #endif
option = 2;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
assert(ghostOwner != -1);
assert(ghostOwner != myRank);
#pragma omp atomic
PCounter[ghostOwner]++;
#pragma omp atomic
(*msgIndPtr)++;
#pragma omp atomic
(*NumMessagesBundledPtr)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(REQUEST);
privateQOwner.push_back(ghostOwner);
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) if (candidateMate[NLVer + Ghost2LocalMap[w]] == v)
{ {
option = 1;
Mate[v - StartIndex] = w; // v is a local vertex Mate[v - StartIndex] = w; // v is a local vertex
GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex
privateU.push_back(v);
privateU.push_back(w);
#pragma omp atomic
(*myCardPtr)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
// Decrement the counter: // Decrement the counter:
PROCESS_CROSS_EDGE(Counter, Ghost2LocalMap[w], SPtr); PROCESS_CROSS_EDGE(Counter, Ghost2LocalMap[w], SPtr);
} // End of if CandidateMate[w] = v } // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex } // End of if a Ghost Vertex
else else
{ // w is a local vertex { // w is a local vertex
if (candidateMate[w - StartIndex] == v) if (candidateMate[w - StartIndex] == v)
{ {
option = 3;
Mate[v - StartIndex] = w; // v is a local vertex Mate[v - StartIndex] = w; // v is a local vertex
Mate[w - StartIndex] = v; // w is a local vertex Mate[w - StartIndex] = v; // w is a local vertex
privateU.push_back(v);
privateU.push_back(w);
#pragma omp atomic
(*myCardPtr)++;
#ifdef PRINT_DEBUG_INFO_ #ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") "; cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout); fflush(stdout);
#endif #endif
} // End of if(CandidateMate(w) = v } // End of if(CandidateMate(w) = v
} // End of Else } // End of Else
} // End of if(w >=0) } // End of if(w >=0)
else else option = 4;// End of Else: w == -1
{ // End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
adj11 = verLocPtr[v - StartIndex]; }
adj12 = verLocPtr[v - StartIndex + 1]; } // End of task
for (k1 = adj11; k1 < adj12; k1++) } // End of If (candidateMate[v-StartIndex] == u
{
w = verLocInd[k1];
if ((w < StartIndex) || (w > EndIndex))
{ // A ghost
} // End of if ( (v >= StartIndex) && (v <= EndIndex) ) //If Local Vertex:
else
{ // Neighbor is a ghost vertex
#pragma omp critical
{
if (candidateMate[NLVer + Ghost2LocalMap[v]] == u)
candidateMate[NLVer + Ghost2LocalMap[v]] = -1;
if (v != Mate[u - StartIndex]) option = 5; // u is local
} // End of critical
} // End of Else //A Ghost Vertex
switch (option)
{
case -1:
// No things to do
break;
case 1:
// Found a dominating edge, it is a ghost and candidateMate[NLVer + Ghost2LocalMap[w]] == v
privateU.push_back(v);
privateU.push_back(w);
#pragma omp atomic
(*myCardPtr)++;
#ifdef PRINT_DEBUG_INFO_ #ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a failure message: "; cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs); fflush(stdout);
fflush(stdout);
#endif #endif
case 2:
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs); // Found a dominating edge, it is a ghost
assert(ghostOwner != -1); ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
assert(ghostOwner != myRank); assert(ghostOwner != -1);
assert(ghostOwner != myRank);
#pragma omp atomic
PCounter[ghostOwner]++;
#pragma omp atomic #pragma omp atomic
PCounter[ghostOwner]++; (*msgIndPtr)++;
#pragma omp atomic #pragma omp atomic
(*msgIndPtr)++; (*NumMessagesBundledPtr)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(REQUEST);
privateQOwner.push_back(ghostOwner);
break;
case 3:
privateU.push_back(v);
privateU.push_back(w);
#pragma omp atomic #pragma omp atomic
(*NumMessagesBundledPtr)++; (*myCardPtr)++;
break;
case 4:
// Could not find a dominating vertex
adj11 = verLocPtr[v - StartIndex];
adj12 = verLocPtr[v - StartIndex + 1];
for (k1 = adj11; k1 < adj12; k1++)
{
w = verLocInd[k1];
if ((w < StartIndex) || (w > EndIndex))
{ // A ghost
privateQLocalVtx.push_back(v); #ifdef PRINT_DEBUG_INFO_
privateQGhostVtx.push_back(w); cout << "\n(" << myRank << ")Sending a failure message: ";
privateQMsgType.push_back(FAILURE); cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
privateQOwner.push_back(ghostOwner); fflush(stdout);
#endif
} // End of if(GHOST) ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
} // End of for loop assert(ghostOwner != -1);
} // End of Else: w == -1 assert(ghostOwner != myRank);
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v) #pragma omp atomic
} PCounter[ghostOwner]++;
} // End of task #pragma omp atomic
} // End of If (candidateMate[v-StartIndex] == u (*msgIndPtr)++;
#pragma omp atomic
(*NumMessagesBundledPtr)++;
} // End of if ( (v >= StartIndex) && (v <= EndIndex) ) //If Local Vertex: privateQLocalVtx.push_back(v);
else privateQGhostVtx.push_back(w);
{ // Neighbor is a ghost vertex privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
#pragma omp critical } // End of if(GHOST)
{ } // End of for loop
if (candidateMate[NLVer + Ghost2LocalMap[v]] == u) break;
candidateMate[NLVer + Ghost2LocalMap[v]] = -1; default:
if (v != Mate[u - StartIndex])
{ // u is local
#ifdef PRINT_DEBUG_INFO_ #ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a success message: "; cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n"; cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n";
fflush(stdout); fflush(stdout);
#endif #endif
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs); ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
assert(ghostOwner != -1); assert(ghostOwner != -1);
assert(ghostOwner != myRank); assert(ghostOwner != myRank);
#pragma omp atomic #pragma omp atomic
PCounter[ghostOwner]++; PCounter[ghostOwner]++;
#pragma omp atomic #pragma omp atomic
(*msgIndPtr)++; (*msgIndPtr)++;
#pragma omp atomic #pragma omp atomic
(*NumMessagesBundledPtr)++; (*NumMessagesBundledPtr)++;
privateQLocalVtx.push_back(u); privateQLocalVtx.push_back(u);
privateQGhostVtx.push_back(v); privateQGhostVtx.push_back(v);
privateQMsgType.push_back(SUCCESS); privateQMsgType.push_back(SUCCESS);
privateQOwner.push_back(ghostOwner); privateQOwner.push_back(ghostOwner);
} // End of If( v != Mate[u] ) break;
} //End of switch
} // End of task } // End of inner for
} // End of Else //A Ghost Vertex
} // End of inner for
// TODO privateU.size() < UCHUNK could be commented but it generate errors, why? // TODO privateU.size() < UCHUNK could be commented but it generate errors, why?
if (privateU.size() > UCHUNK || U.empty()) if (privateU.size() > UCHUNK || U.empty())
{ {
#pragma omp critical(U) #pragma omp critical(U)
{ {
while (!privateU.empty()) while (!privateU.empty())
U.push_back(privateU.pop_back()); U.push_back(privateU.pop_back());
} }
#ifndef error #ifndef error
#pragma omp critical(privateMsg) #pragma omp critical(privateMsg)
{
while (!privateQLocalVtx.empty())
{ {
while (!privateQLocalVtx.empty()) QLocalVtx.push_back(privateQLocalVtx.pop_back());
{ QGhostVtx.push_back(privateQGhostVtx.pop_back());
QLocalVtx.push_back(privateQLocalVtx.pop_back()); QMsgType.push_back(privateQMsgType.pop_back());
QGhostVtx.push_back(privateQGhostVtx.pop_back()); QOwner.push_back(privateQOwner.pop_back());
QMsgType.push_back(privateQMsgType.pop_back());
QOwner.push_back(privateQOwner.pop_back());
}
} }
}
#endif #endif
} // End of private.size() } // End of private.size()
} }
} // End of outer for } // End of outer for
} // End of while ( !U.empty() ) } // End of while ( !U.empty() )
queuesTransfer(U, privateU, QLocalVtx, queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx, QGhostVtx,
QMsgType, QOwner, privateQLocalVtx, QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx, privateQGhostVtx,
privateQMsgType, privateQMsgType,
privateQOwner); privateQOwner);
#ifdef COUNT_LOCAL_VERTEX #ifdef COUNT_LOCAL_VERTEX
printf("Count local vertexes: %ld for thread %d of processor %d\n", printf("Count local vertexes: %ld for thread %d of processor %d\n",
localVertices, localVertices,
omp_get_thread_num(), omp_get_thread_num(),
myRank); myRank);
#endif #endif
} // End of parallel region } // End of parallel region

Write Preview
Loading…
Cancel
Save