fdurastante/amg4psblas
1
0
Fork 0
mirror of https://github.com/sfilippone/amg4psblas.git synced 2026-10-06 14:44:55 +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/algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP.cpp
+122 -53
View File
@@ -70,7 +70,7 @@
Statistics: ph0_time, ph1_time, ph2_time: Runtimes
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
#else
@@ -110,17 +110,24 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ") verDistance [" << verDistance[0] << "," << verDistance[1] << "," << verDistance[2] << "," << verDistance[3] << "]";
cout << "\n(" << myRank << ") verDistance [" ;
for (int i = 0; i < numProcs; i++)
cout << verDistance[i] << "," << verDistance[i+1];
cout << "]\n";
fflush(stdout);
#endif
#ifdef DEBUG_HANG_
if (myRank == 0)
cout << "\n(" << myRank << ") verDistance [" << verDistance[0] << "," << verDistance[1] << "," << verDistance[2] << "," << verDistance[3] << "]";
if (myRank == 0) {
cout << "\n(" << myRank << ") verDistance [" ;
for (int i = 0; i < numProcs; i++)
cout << verDistance[i] << "," ;
cout << verDistance[numProcs]<< "]\n";
}
fflush(stdout);
#endif
MilanLongInt StartIndex = verDistance[myRank]; // The starting vertex owned by the current rank
MilanLongInt EndIndex = verDistance[myRank + 1] - 1; // The ending vertex owned by the current rank
MilanLongInt EndIndex = verDistance[myRank + 1] - 1; // The ending vertex owned by the current rank
MPI_Status computeStatus;
@@ -211,7 +218,11 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
finishTime = MPI_Wtime();
*ph0_time = finishTime - startTime; // Time taken for Phase-0: Initialization
#ifdef DEBUG_HANG_
cout << myRank << " Finished initialization" << endl;
fflush(stdout);
#endif
startTime = MPI_Wtime();
/////////////////////////////////////////////////////////////////////////////////////////
@@ -233,6 +244,17 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
edgeLocWeight,
candidateMate);
#ifdef DEBUG_HANG_
cout << myRank << " Finished Exposed Vertex" << endl;
fflush(stdout);
#if 0
cout << myRank << " candidateMate after parallelCompute " <<endl;
for (int i=0; i<NLVer; i++) {
cout << candidateMate[i] << " " ;
}
cout << endl;
#endif
#endif
/*
* PARALLEL_PROCESS_EXPOSED_VERTEX_B
* TODO: write comment
@@ -272,6 +294,18 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
tempCounter.clear(); // Do not need this any more
#ifdef DEBUG_HANG_
cout << myRank << " Finished Exposed Vertex" << endl;
fflush(stdout);
#if 0
cout << myRank << " Mate after Exposed Vertices " <<endl;
for (int i=0; i<NLVer; i++) {
cout << Mate[i] << " " ;
}
cout << endl;
#endif
#endif
///////////////////////////////////////////////////////////////////////////////////
/////////////////////////// PROCESS MATCHED VERTICES //////////////////////////////
///////////////////////////////////////////////////////////////////////////////////
@@ -311,6 +345,19 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
privateQMsgType,
privateQOwner);
#ifdef DEBUG_HANG_
cout << myRank << " Finished Process Vertices" << endl;
fflush(stdout);
#if 0
cout << myRank << " Mate after Matched Vertices " <<endl;
for (int i=0; i<NLVer; i++) {
cout << Mate[i] << " " ;
}
cout << endl;
#endif
#endif
/////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////// SEND BUNDLED MESSAGES /////////////////////////////////////
/////////////////////////////////////////////////////////////////////////////////////////
@@ -339,6 +386,12 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
finishTime = MPI_Wtime();
*ph1_time = finishTime - startTime; // Time taken for Phase-1
#ifdef DEBUG_HANG_
cout << myRank << " Finished sendBundles" << endl;
fflush(stdout);
#endif
*ph1_card = myCard; // Cardinality at the end of Phase-1
startTime = MPI_Wtime();
/////////////////////////////////////////////////////////////////////////////////////////
@@ -360,65 +413,72 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
fflush(stdout);
#endif
while (true)
{
while (true) {
#ifdef DEBUG_HANG_
if (myRank == 0)
cout << "\n(" << myRank << ") Main loop" << endl;
fflush(stdout);
//if (myRank == 0)
cout << "\n(" << myRank << ") Main loop" << endl;
fflush(stdout);
#endif
///////////////////////////////////////////////////////////////////////////////////
/////////////////////////// PROCESS MATCHED VERTICES //////////////////////////////
///////////////////////////////////////////////////////////////////////////////////
processMatchedVerticesAndSendMessages(NLVer,
UChunkBeingProcessed,
U,
privateU,
StartIndex,
EndIndex,
&myCard,
&msgInd,
&NumMessagesBundled,
&S,
verLocPtr,
verLocInd,
verDistance,
PCounter,
Counter,
myRank,
numProcs,
candidateMate,
GMate,
Mate,
Ghost2LocalMap,
edgeLocWeight,
QLocalVtx,
QGhostVtx,
QMsgType,
QOwner,
privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner,
comm,
&msgActual,
Message);
///////////////////////// END OF PROCESS MATCHED VERTICES /////////////////////////
processMatchedVerticesAndSendMessages(NLVer,
UChunkBeingProcessed,
U,
privateU,
StartIndex,
EndIndex,
&myCard,
&msgInd,
&NumMessagesBundled,
&S,
verLocPtr,
verLocInd,
verDistance,
PCounter,
Counter,
myRank,
numProcs,
candidateMate,
GMate,
Mate,
Ghost2LocalMap,
edgeLocWeight,
QLocalVtx,
QGhostVtx,
QMsgType,
QOwner,
privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner,
comm,
&msgActual,
Message);
///////////////////////// END OF PROCESS MATCHED VERTICES /////////////////////////
//// BREAK IF NO MESSAGES EXPECTED /////////
#ifdef DEBUG_HANG_
#if 0
cout << myRank << " Mate after ProcessMatchedAndSend phase "<<S <<endl;
for (int i=0; i<NLVer; i++) {
cout << Mate[i] << " " ;
}
cout << endl;
#endif
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Deciding whether to break: S= " << S << endl;
#endif
if (S == 0)
{
if (S == 0) {
#ifdef DEBUG_HANG_
cout << "\n(" << myRank << ") Breaking out" << endl;
fflush(stdout);
cout << "\n(" << myRank << ") Breaking out" << endl;
fflush(stdout);
#endif
break;
break;
}
///////////////////////////////////////////////////////////////////////////////////
/////////////////////////// PROCESS MESSAGES //////////////////////////////////////
@@ -451,6 +511,15 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
U);
///////////////////////// END OF PROCESS MESSAGES /////////////////////////////////
#ifdef DEBUG_HANG_
#if 0
cout << myRank << " Mate after ProcessMessages phase "<<S <<endl;
for (int i=0; i<NLVer; i++) {
cout << Mate[i] << " " ;
}
cout << endl;
#endif
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Finished Message processing phase: S= " << S;
fflush(stdout);
@@ -482,4 +551,4 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
// End of algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMate
#endif
#endif
#endif
amgprec/impl/aggregator/computeCandidateMate.cpp
+18 -23
View File
@@ -17,16 +17,13 @@ MilanLongInt firstComputeCandidateMate(MilanLongInt adj1,
MilanInt w = -1;
MilanReal heaviestEdgeWt = MilanRealMin; // Assign the smallest Value possible first LDBL_MIN
int finalK;
for (int k = adj1; k < adj2; k++)
{
if ((edgeLocWeight[k] > heaviestEdgeWt) ||
((edgeLocWeight[k] == heaviestEdgeWt) && (w < verLocInd[k])))
{
heaviestEdgeWt = edgeLocWeight[k];
w = verLocInd[k];
finalK = k;
}
for (int k = adj1; k < adj2; k++) {
if ((edgeLocWeight[k] > heaviestEdgeWt) ||
((edgeLocWeight[k] == heaviestEdgeWt) && (w < verLocInd[k]))) {
heaviestEdgeWt = edgeLocWeight[k];
w = verLocInd[k];
finalK = k;
}
} // End of for loop
return finalK;
}
@@ -60,19 +57,17 @@ MilanLongInt computeCandidateMate(MilanLongInt adj1,
MilanInt w = -1;
MilanReal heaviestEdgeWt = MilanRealMin; // Assign the smallest Value possible first LDBL_MIN
for (k = adj1; k < adj2; k++)
{
if (isAlreadyMatched(verLocInd[k], StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap))
continue;
if ((edgeLocWeight[k] > heaviestEdgeWt) ||
((edgeLocWeight[k] == heaviestEdgeWt) && (w < verLocInd[k])))
{
heaviestEdgeWt = edgeLocWeight[k];
w = verLocInd[k];
}
for (k = adj1; k < adj2; k++) {
if (isAlreadyMatched(verLocInd[k], StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap))
continue;
if ((edgeLocWeight[k] > heaviestEdgeWt) ||
((edgeLocWeight[k] == heaviestEdgeWt) && (w < verLocInd[k]))) {
heaviestEdgeWt = edgeLocWeight[k];
w = verLocInd[k];
}
} // End of for loop
// End: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
return w;
}
}
amgprec/impl/aggregator/parallelComputeCandidateMateB.cpp
+1 -2
View File
@@ -14,8 +14,7 @@ void PARALLEL_COMPUTE_CANDIDATE_MATE_B(MilanLongInt NLVer,
{
#pragma omp for schedule(static)
for (v = 0; v < NLVer; v++)
{
for (v = 0; v < NLVer; v++) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Processing: " << v + StartIndex << endl;
fflush(stdout);
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
}
}
amgprec/impl/aggregator/processMatchedVertices.cpp
+211 -226
View File
@@ -46,264 +46,249 @@ void processMatchedVertices(
#ifdef COUNT_LOCAL_VERTEX
MilanLongInt localVertices = 0;
#endif
#pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner, option) \
firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx, privateQMsgType, privateQOwner, UChunkBeingProcessed) default(shared) \
num_threads(NUM_THREAD) \
reduction(+ \
: msgInd[:1], PCounter \
[:numProcs], myCard \
[:1], NumMessagesBundled \
//#pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner, option) \
firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx, \
privateQMsgType, privateQOwner, UChunkBeingProcessed) \
default(shared) num_threads(NUM_THREAD) \
reduction(+ \
: msgInd[:1], PCounter \
[:numProcs], myCard \
[:1], NumMessagesBundled \
[:1])
{
while (!U.empty())
{
extractUChunk(UChunkBeingProcessed, U, privateU);
for (MilanLongInt u : UChunkBeingProcessed)
{
while (!U.empty()) {
extractUChunk(UChunkBeingProcessed, U, privateU);
for (MilanLongInt u : UChunkBeingProcessed) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")u: " << u;
fflush(stdout);
cout << "\n(" << myRank << ")u: " << u;
fflush(stdout);
#endif
if ((u >= StartIndex) && (u <= EndIndex))
{ // Process Only the Local Vertices
if ((u >= StartIndex) && (u <= EndIndex)) { // Process Only the Local Vertices
#ifdef COUNT_LOCAL_VERTEX
localVertices++;
localVertices++;
#endif
// Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++)
{
option = -1;
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex))
{ // If Local Vertex:
// Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++) {
option = -1;
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex)) { // If Local Vertex:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout);
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout);
#endif
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0)
{
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
#pragma omp critical
{
if (candidateMate[v - StartIndex] == u)
{
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMate(verLocPtr[v - StartIndex],
verLocPtr[v - StartIndex + 1],
edgeLocWeight, 0,
verLocInd,
StartIndex,
EndIndex,
GMate,
Mate,
Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
{
if (candidateMate[v - StartIndex] == u) {
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMate(verLocPtr[v - StartIndex],
verLocPtr[v - StartIndex + 1],
edgeLocWeight, 0,
verLocInd,
StartIndex,
EndIndex,
GMate,
Mate,
Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w;
fflush(stdout);
cout << "\n(" << myRank << ")" << v << " Points to: " << w;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0)
{
if ((w < StartIndex) || (w > EndIndex))
{ // A ghost
// If found a dominating edge:
if (w >= 0) {
if ((w < StartIndex) || (w > EndIndex)) { // A ghost
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message:";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
cout << "\n(" << myRank << ")Sending a request message:";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
#endif
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v)
{
option = 1;
Mate[v - StartIndex] = w; // v is a local vertex
GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else
{ // w is a local vertex
if (candidateMate[w - StartIndex] == v)
{
option = 3;
Mate[v - StartIndex] = w; // v is a local vertex
Mate[w - StartIndex] = v; // w is a local vertex
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) {
option = 1;
Mate[v - StartIndex] = w; // v is a local vertex
GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == v) {
option = 3;
Mate[v - StartIndex] = w; // v is a local vertex
Mate[w - StartIndex] = v; // w is a local vertex
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
option = 4; // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of If (candidateMate[v-StartIndex] == u
} // End of task
} // mateval < 0
} // End of if ( (v >= StartIndex) && (v <= EndIndex) ) //If Local Vertex:
else
{ // Neighbor is a ghost vertex
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
option = 4; // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of If (candidateMate[v-StartIndex] == u
} // End of task
} // mateval < 0
} // 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);
(*myCard)++;
{
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);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], SPtr);
case 2:
// Found a dominating edge, it is a ghost
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
PCounter[ghostOwner]++;
(*NumMessagesBundled)++;
(*msgInd)++;
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);
(*myCard)++;
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
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], SPtr);
case 2:
// Found a dominating edge, it is a ghost
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
PCounter[ghostOwner]++;
(*NumMessagesBundled)++;
(*msgInd)++;
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);
(*myCard)++;
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
#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
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
PCounter[ghostOwner]++;
(*NumMessagesBundled)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
break;
case 5:
default:
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
PCounter[ghostOwner]++;
(*NumMessagesBundled)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
break;
case 5:
default:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n";
fflush(stdout);
cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n";
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
(*NumMessagesBundled)++;
PCounter[ghostOwner]++;
(*msgInd)++;
privateQLocalVtx.push_back(u);
privateQGhostVtx.push_back(v);
privateQMsgType.push_back(SUCCESS);
privateQOwner.push_back(ghostOwner);
break;
} // End of switch
} // End of inner for
}
} // End of outer for
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
(*NumMessagesBundled)++;
PCounter[ghostOwner]++;
(*msgInd)++;
privateQLocalVtx.push_back(u);
privateQGhostVtx.push_back(v);
privateQMsgType.push_back(SUCCESS);
privateQOwner.push_back(ghostOwner);
break;
} // End of switch
} // End of inner for
}
} // End of outer for
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
#pragma omp critical(U)
{
U.insert(U.end(), privateU.begin(), privateU.end());
}
privateU.clear();
{
U.insert(U.end(), privateU.begin(), privateU.end());
}
privateU.clear();
#pragma omp critical(sendMessageTransfer)
{
QLocalVtx.insert(QLocalVtx.end(), privateQLocalVtx.begin(), privateQLocalVtx.end());
QGhostVtx.insert(QGhostVtx.end(), privateQGhostVtx.begin(), privateQGhostVtx.end());
QMsgType.insert(QMsgType.end(), privateQMsgType.begin(), privateQMsgType.end());
QOwner.insert(QOwner.end(), privateQOwner.begin(), privateQOwner.end());
}
privateQLocalVtx.clear();
privateQGhostVtx.clear();
privateQMsgType.clear();
privateQOwner.clear();
{
QLocalVtx.insert(QLocalVtx.end(), privateQLocalVtx.begin(), privateQLocalVtx.end());
QGhostVtx.insert(QGhostVtx.end(), privateQGhostVtx.begin(), privateQGhostVtx.end());
QMsgType.insert(QMsgType.end(), privateQMsgType.begin(), privateQMsgType.end());
QOwner.insert(QOwner.end(), privateQOwner.begin(), privateQOwner.end());
}
privateQLocalVtx.clear();
privateQGhostVtx.clear();
privateQMsgType.clear();
privateQOwner.clear();
} // End of while ( !U.empty() )
#ifdef COUNT_LOCAL_VERTEX
printf("Count local vertexes: %ld for thread %d of processor %d\n",
localVertices,
omp_get_thread_num(),
myRank);
#endif
} // End of parallel region
}
amgprec/impl/aggregator/processMatchedVerticesAndSendMessages.cpp
+255 -262
View File
@@ -1,39 +1,39 @@
#include "MatchBoxPC.h"
//#define DEBUG_HANG_
void processMatchedVerticesAndSendMessages(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *NumMessagesBundled,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner,
MPI_Comm comm,
MilanLongInt *msgActual,
vector<MilanLongInt> &Message)
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *NumMessagesBundled,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner,
MPI_Comm comm,
MilanLongInt *msgActual,
vector<MilanLongInt> &Message)
{
MilanLongInt initialSize = QLocalVtx.size();
@@ -50,266 +50,259 @@ void processMatchedVerticesAndSendMessages(
#ifdef COUNT_LOCAL_VERTEX
MilanLongInt localVertices = 0;
#endif
#pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner, option) \
firstprivate(Message, privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx, privateQMsgType, privateQOwner, UChunkBeingProcessed) default(shared) \
num_threads(NUM_THREAD) \
reduction(+ \
: msgInd[:1], PCounter \
[:numProcs], myCard \
[:1], NumMessagesBundled \
[:1], msgActual \
//#pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner, option) \
firstprivate(Message, privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx,\
privateQMsgType, privateQOwner, UChunkBeingProcessed) default(shared) \
num_threads(NUM_THREAD) \
reduction(+ \
: msgInd[:1], PCounter \
[:numProcs], myCard \
[:1], NumMessagesBundled \
[:1], msgActual \
[:1])
{
while (!U.empty())
{
extractUChunk(UChunkBeingProcessed, U, privateU);
for (MilanLongInt u : UChunkBeingProcessed)
{
while (!U.empty()) {
extractUChunk(UChunkBeingProcessed, U, privateU);
for (MilanLongInt u : UChunkBeingProcessed) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")u: " << u;
fflush(stdout);
cout << "\n(" << myRank << ")u: " << u;
fflush(stdout);
#endif
if ((u >= StartIndex) && (u <= EndIndex))
{ // Process Only the Local Vertices
if ((u >= StartIndex) && (u <= EndIndex)) { // Process Only the Local Vertices
#ifdef COUNT_LOCAL_VERTEX
localVertices++;
localVertices++;
#endif
// Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++)
{
option = -1;
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex))
{ // If Local Vertex:
// Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++) {
option = -1;
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex)) { // If Local Vertex:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout);
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout);
#endif
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0)
{
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
#pragma omp critical
{
if (candidateMate[v - StartIndex] == u)
{
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMate(verLocPtr[v - StartIndex],
verLocPtr[v - StartIndex + 1],
edgeLocWeight, 0,
verLocInd,
StartIndex,
EndIndex,
GMate,
Mate,
Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
{
if (candidateMate[v - StartIndex] == u) {
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMate(verLocPtr[v - StartIndex],
verLocPtr[v - StartIndex + 1],
edgeLocWeight, 0,
verLocInd,
StartIndex,
EndIndex,
GMate,
Mate,
Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w;
fflush(stdout);
cout << "\n(" << myRank << ")" << v << " Points to: " << w;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0)
{
if ((w < StartIndex) || (w > EndIndex))
{ // A ghost
// If found a dominating edge:
if (w >= 0) {
if ((w < StartIndex) || (w > EndIndex)) { // A ghost
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message:";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
cout << "\n(" << myRank << ")Sending a request message:";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
#endif
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v)
{
option = 1;
Mate[v - StartIndex] = w; // v is a local vertex
GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else
{ // w is a local vertex
if (candidateMate[w - StartIndex] == v)
{
option = 3;
Mate[v - StartIndex] = w; // v is a local vertex
Mate[w - StartIndex] = v; // w is a local vertex
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) {
option = 1;
Mate[v - StartIndex] = w; // v is a local vertex
GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == v) {
option = 3;
Mate[v - StartIndex] = w; // v is a local vertex
Mate[w - StartIndex] = v; // w is a local vertex
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
option = 4; // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of If (candidateMate[v-StartIndex] == u
} // End of task
} // mateval < 0
} // End of if ( (v >= StartIndex) && (v <= EndIndex) ) //If Local Vertex:
else
{ // Neighbor is a ghost vertex
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
option = 4; // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of If (candidateMate[v-StartIndex] == u
} // End of task
} // mateval < 0
} // 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);
(*myCard)++;
{
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);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], SPtr);
case 2:
// Found a dominating edge, it is a ghost
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = v; // LOCAL
// Message[1] = w; // GHOST
// Message[2] = REQUEST; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
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);
(*myCard)++;
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
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], SPtr);
case 2:
// Found a dominating edge, it is a ghost
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = v; // LOCAL
// Message[1] = w; // GHOST
// Message[2] = REQUEST; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
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);
(*myCard)++;
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
#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
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = v; // LOCAL
// Message[1] = w; // GHOST
// Message[2] = FAILURE; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
break;
case 5:
default:
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = v; // LOCAL
// Message[1] = w; // GHOST
// Message[2] = FAILURE; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
break;
case 5:
default:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n";
fflush(stdout);
cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n";
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = u; // LOCAL
// Message[1] = v; // GHOST
// Message[2] = SUCCESS; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(u);
privateQGhostVtx.push_back(v);
privateQMsgType.push_back(SUCCESS);
privateQOwner.push_back(ghostOwner);
break;
} // End of switch
} // End of inner for
}
} // End of outer for
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = u; // LOCAL
// Message[1] = v; // GHOST
// Message[2] = SUCCESS; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(u);
privateQGhostVtx.push_back(v);
privateQMsgType.push_back(SUCCESS);
privateQOwner.push_back(ghostOwner);
break;
} // End of switch
} // End of inner for
}
} // End of outer for
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
} // End of while ( !U.empty() )
#ifdef COUNT_LOCAL_VERTEX
printf("Count local vertexes: %ld for thread %d of processor %d\n",
localVertices,
omp_get_thread_num(),
myRank);
#endif
} // End of parallel region
// Send the messages
for (int i = initialSize; i < QOwner.size(); i++)
{
Message[0] = QLocalVtx[i];
Message[1] = QGhostVtx[i];
Message[2] = QMsgType[i];
ghostOwner = QOwner[i];
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
#ifdef DEBUG_HANG_
cout << myRank<<" Sending: "<<QOwner.size()-initialSize<<" messages" <<endl;
#endif
for (int i = initialSize; i < QOwner.size(); i++) {
Message[0] = QLocalVtx[i];
Message[1] = QGhostVtx[i];
Message[2] = QMsgType[i];
ghostOwner = QOwner[i];
//MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
//cout << myRank<<" Sending to "<<ghostOwner<<endl;
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
}
#ifdef DEBUG_HANG_
cout << myRank<<" Done sending messages"<<endl;
#endif
}
amgprec/impl/aggregator/processMessages.cpp
+178 -192
View File
@@ -1,4 +1,5 @@
#include "MatchBoxPC.h"
//#define DEBUG_HANG_
void processMessages(
MilanLongInt NLVer,
@@ -78,6 +79,7 @@ void processMessages(
fflush(stdout);
#endif
//cout << myRank<<" Receiving ...";
error_codeC = MPI_Recv(&Message[0], 3, TypeMap<MilanLongInt>(), MPI_ANY_SOURCE, ComputeTag, comm, &computeStatus);
if (error_codeC != MPI_SUCCESS)
{
@@ -86,70 +88,66 @@ void processMessages(
fflush(stdout);
}
Sender = computeStatus.MPI_SOURCE;
//cout << " ...from "<<Sender << endl;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received message from Process " << Sender << " Type= " << Message[2] << endl;
fflush(stdout);
#endif
if (Message[2] == SIZEINFO)
{
if (Message[2] == SIZEINFO) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received bundled message from Process " << Sender << " Size= " << Message[0] << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Received bundled message from Process " << Sender << " Size= " << Message[0] << endl;
fflush(stdout);
#endif
bundleSize = Message[0]; //#of integers in the message
// Build the Message Buffer:
if (!ReceiveBuffer.empty())
ReceiveBuffer.clear(); // Empty it out first
ReceiveBuffer.resize(bundleSize, -1); // Initialize
bundleSize = Message[0]; //#of integers in the message
// Build the Message Buffer:
if (!ReceiveBuffer.empty())
ReceiveBuffer.clear(); // Empty it out first
ReceiveBuffer.resize(bundleSize, -1); // Initialize
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message Bundle Before: " << endl;
for (int i = 0; i < bundleSize; i++)
cout << ReceiveBuffer[i] << ",";
cout << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Message Bundle Before: " << endl;
for (int i = 0; i < bundleSize; i++)
cout << ReceiveBuffer[i] << ",";
cout << endl;
fflush(stdout);
#endif
// Receive the message
error_codeC = MPI_Recv(&ReceiveBuffer[0], bundleSize, TypeMap<MilanLongInt>(), Sender, BundleTag, comm, &computeStatus);
if (error_codeC != MPI_SUCCESS)
{
MPI_Error_string(error_codeC, error_message, &message_length);
cout << "\n*Error in call to MPI_Receive on processor " << myRank << " Error: " << error_message << "\n";
fflush(stdout);
}
// Receive the message
//cout << myRank<<" Receiving from "<<Sender<<endl;
error_codeC = MPI_Recv(&ReceiveBuffer[0], bundleSize, TypeMap<MilanLongInt>(), Sender, BundleTag, comm, &computeStatus);
if (error_codeC != MPI_SUCCESS) {
MPI_Error_string(error_codeC, error_message, &message_length);
cout << "\n*Error in call to MPI_Receive on processor " << myRank << " Error: " << error_message << "\n";
fflush(stdout);
}
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message Bundle After: " << endl;
for (int i = 0; i < bundleSize; i++)
cout << ReceiveBuffer[i] << ",";
cout << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Message Bundle After: " << endl;
for (int i = 0; i < bundleSize; i++)
cout << ReceiveBuffer[i] << ",";
cout << endl;
fflush(stdout);
#endif
} else { // Just a single message:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received regular message from Process " << Sender << " u= " << Message[0] << " v= " << Message[1] << endl;
fflush(stdout);
#endif
// Add the current message to Queue:
bundleSize = 3; //#of integers in the message
// Build the Message Buffer:
if (!ReceiveBuffer.empty())
ReceiveBuffer.clear(); // Empty it out first
ReceiveBuffer.resize(bundleSize, -1); // Initialize
ReceiveBuffer[0] = Message[0]; // u
ReceiveBuffer[1] = Message[1]; // v
ReceiveBuffer[2] = Message[2]; // message_type
}
else
{ // Just a single message:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received regular message from Process " << Sender << " u= " << Message[0] << " v= " << Message[1] << endl;
fflush(stdout);
#endif
// Add the current message to Queue:
bundleSize = 3; //#of integers in the message
// Build the Message Buffer:
if (!ReceiveBuffer.empty())
ReceiveBuffer.clear(); // Empty it out first
ReceiveBuffer.resize(bundleSize, -1); // Initialize
ReceiveBuffer[0] = Message[0]; // u
ReceiveBuffer[1] = Message[1]; // v
ReceiveBuffer[2] = Message[2]; // message_type
}
#ifdef DEBUG_GHOST_
if ((v < StartIndex) || (v > EndIndex))
{
cout << "\n(" << myRank << ") From ReceiveBuffer: This should not happen: u= " << u << " v= " << v << " Type= " << message_type << " StartIndex " << StartIndex << " EndIndex " << EndIndex << endl;
fflush(stdout);
if ((v < StartIndex) || (v > EndIndex)) {
cout << "\n(" << myRank << ") From ReceiveBuffer: This should not happen: u= " << u << " v= " << v << " Type= " << message_type << " StartIndex " << StartIndex << " EndIndex " << EndIndex << endl;
fflush(stdout);
}
#endif
#ifdef PRINT_DEBUG_INFO_
@@ -158,172 +156,160 @@ void processMessages(
#endif
// Most of the time bundleSize == 3, thus, it's not worth parallelizing thi loop
for (MilanLongInt bundleCounter = 3; bundleCounter < bundleSize + 3; bundleCounter += 3)
{
u = ReceiveBuffer[bundleCounter - 3]; // GHOST
v = ReceiveBuffer[bundleCounter - 2]; // LOCAL
message_type = ReceiveBuffer[bundleCounter - 1]; // TYPE
// CASE I: REQUEST
if (message_type == REQUEST)
{
for (MilanLongInt bundleCounter = 3; bundleCounter < bundleSize + 3; bundleCounter += 3) {
u = ReceiveBuffer[bundleCounter - 3]; // GHOST
v = ReceiveBuffer[bundleCounter - 2]; // LOCAL
message_type = ReceiveBuffer[bundleCounter - 1]; // TYPE
// CASE I: REQUEST
if (message_type == REQUEST) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message type is REQUEST" << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Message type is REQUEST" << endl;
fflush(stdout);
#endif
#ifdef DEBUG_GHOST_
if ((v < 0) || (v < StartIndex) || ((v - StartIndex) > NLVer))
{
cout << "\n(" << myRank << ") case 1 Bad address " << v << " " << StartIndex << " " << v - StartIndex << " " << NLVer << endl;
fflush(stdout);
}
if ((v < 0) || (v < StartIndex) || ((v - StartIndex) > NLVer)) {
cout << "\n(" << myRank << ") case 1 Bad address " << v << " " << StartIndex << " " << v - StartIndex << " " << NLVer << endl;
fflush(stdout);
}
#endif
if (Mate[v - StartIndex] == -1)
{ // Process only if not already matched (v is local)
candidateMate[NLVer + Ghost2LocalMap[u]] = v; // Set CandidateMate for the ghost
if (candidateMate[v - StartIndex] == u)
{
GMate[Ghost2LocalMap[u]] = v; // u is ghost
Mate[v - StartIndex] = u; // v is local
U.push_back(v);
U.push_back(u);
(*myCard)++;
if (Mate[v - StartIndex] == -1) {
// Process only if not already matched (v is local)
candidateMate[NLVer + Ghost2LocalMap[u]] = v; // Set CandidateMate for the ghost
if (candidateMate[v - StartIndex] == u) {
GMate[Ghost2LocalMap[u]] = v; // u is ghost
Mate[v - StartIndex] = u; // v is local
U.push_back(v);
U.push_back(u);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << u << ") " << endl;
fflush(stdout);
cout << "\n(" << myRank << ")MATCH: (" << v << "," << u << ") " << endl;
fflush(stdout);
#endif
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S);
} // End of if ( candidateMate[v-StartIndex] == u )e
} // End of if ( Mate[v] == -1 )
} // End of REQUEST
else
{ // CASE II: SUCCESS
if (message_type == SUCCESS)
{
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S);
} // End of if ( candidateMate[v-StartIndex] == u )e
} // End of if ( Mate[v] == -1 )
} // End of REQUEST
else { // CASE II: SUCCESS
if (message_type == SUCCESS) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message type is SUCCESS" << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Message type is SUCCESS" << endl;
fflush(stdout);
#endif
GMate[Ghost2LocalMap[u]] = EndIndex + 1; // Set a Dummy Mate to make sure that we do not (u is a ghost) process it again
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S);
GMate[Ghost2LocalMap[u]] = EndIndex + 1; // Set a Dummy Mate to make sure that we do not (u is a ghost) process it again
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S);
#ifdef DEBUG_GHOST_
if ((v < 0) || (v < StartIndex) || ((v - StartIndex) > NLVer))
{
cout << "\n(" << myRank << ") case 2 Bad address " << v << " " << StartIndex << " " << v - StartIndex << " " << NLVer << endl;
fflush(stdout);
}
if ((v < 0) || (v < StartIndex) || ((v - StartIndex) > NLVer)) {
cout << "\n(" << myRank << ") case 2 Bad address " << v << " " << StartIndex << " " << v - StartIndex << " " << NLVer << endl;
fflush(stdout);
}
#endif
if (Mate[v - StartIndex] == -1)
{ // Process only if not already matched ( v is local)
if (candidateMate[v - StartIndex] == u)
{
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMate(verLocPtr[v - StartIndex], verLocPtr[v - StartIndex + 1], edgeLocWeight, k, verLocInd, StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
if (Mate[v - StartIndex] == -1) {
// Process only if not already matched ( v is local)
if (candidateMate[v - StartIndex] == u) {
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMate(verLocPtr[v - StartIndex], verLocPtr[v - StartIndex + 1], edgeLocWeight, k,
verLocInd, StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w << endl;
fflush(stdout);
cout << "\n(" << myRank << ")" << v << " Points to: " << w << endl;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0)
{
if ((w < StartIndex) || (w > EndIndex))
{ // w is a ghost
// Build the Message Packet:
Message[0] = v; // LOCAL
Message[1] = w; // GHOST
Message[2] = REQUEST; // TYPE
// Send a Request (Asynchronous)
// If found a dominating edge:
if (w >= 0) {
if ((w < StartIndex) || (w > EndIndex)) {
// w is a ghost
// Build the Message Packet:
Message[0] = v; // LOCAL
Message[1] = w; // GHOST
Message[2] = REQUEST; // TYPE
// Send a Request (Asynchronous)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Sending a request message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
//assert(ghostOwner != -1);
//assert(ghostOwner != myRank);
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgInd)++;
(*msgActual)++;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v)
{
Mate[v - StartIndex] = w; // v is local
GMate[Ghost2LocalMap[w]] = v; // w is ghost
U.push_back(v);
U.push_back(w);
(*myCard)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
//assert(ghostOwner != -1);
//assert(ghostOwner != myRank);
//cout << myRank<<" Sending to "<<ghostOwner<<endl;
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgInd)++;
(*msgActual)++;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) {
Mate[v - StartIndex] = w; // v is local
GMate[Ghost2LocalMap[w]] = v; // w is ghost
U.push_back(v);
U.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
#endif
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], S);
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else
{ // w is a local vertex
if (candidateMate[w - StartIndex] == v)
{
Mate[v - StartIndex] = w; // v is local
Mate[w - StartIndex] = v; // w is local
// Q.push_back(u);
U.push_back(v);
U.push_back(w);
(*myCard)++;
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], S);
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == v) {
Mate[v - StartIndex] = w; // v is local
Mate[w - StartIndex] = v; // w is local
// Q.push_back(u);
U.push_back(v);
U.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
{ // No dominant edge found
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
// Build the Message Packet:
Message[0] = v; // LOCAL
Message[1] = w; // GHOST
Message[2] = FAILURE; // TYPE
// Send a Request (Asynchronous)
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else { // No dominant edge found
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
// Build the Message Packet:
Message[0] = v; // LOCAL
Message[1] = w; // GHOST
Message[2] = FAILURE; // TYPE
// Send a Request (Asynchronous)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a failure message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Sending a failure message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
//assert(ghostOwner != -1);
//assert(ghostOwner != myRank);
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgInd)++;
(*msgActual)++;
} // End of if(GHOST)
} // End of for loop
} // End of Else: w == -1
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
//assert(ghostOwner != -1);
//assert(ghostOwner != myRank);
//cout << myRank<<" Sending to "<<ghostOwner<<endl;
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgInd)++;
(*msgActual)++;
} // End of if(GHOST)
} // End of for loop
} // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of if ( candidateMate[v-StartIndex] == u )
} // End of if ( Mate[v] == -1 )
} // End of if ( message_type == SUCCESS )
else
{ // CASE III: FAILURE
} // End of if ( candidateMate[v-StartIndex] == u )
} // End of if ( Mate[v] == -1 )
} // End of if ( message_type == SUCCESS )
else {
// CASE III: FAILURE
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message type is FAILURE" << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Message type is FAILURE" << endl;
fflush(stdout);
#endif
GMate[Ghost2LocalMap[u]] = EndIndex + 1; // Set a Dummy Mate to make sure that we do not (u is a ghost) process this anymore
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S); // Decrease the counter
} // End of else: CASE III
} // End of else: CASE I
GMate[Ghost2LocalMap[u]] = EndIndex + 1; // Set a Dummy Mate to make sure that we do not (u is a ghost) process this anymore
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S); // Decrease the counter
} // End of else: CASE III
} // End of else: CASE I
}
return;
}
}
amgprec/impl/aggregator/queueTransfer.cpp
+2 -1
View File
@@ -32,4 +32,5 @@ void queuesTransfer(vector<MilanLongInt> &U,
privateQGhostVtx.clear();
privateQMsgType.clear();
privateQOwner.clear();
}
}
amgprec/impl/aggregator/sendBundledMessages.cpp
+120 -121
View File
@@ -38,108 +38,107 @@ void sendBundledMessages(MilanLongInt *numGhostEdges,
#pragma omp task depend(inout \
: PCumulative, PMessageBundle, PSizeInfoMessages) depend(in \
: NumMessagesBundled, numProcs)
{try {
{
try {
PMessageBundle.reserve(NumMessagesBundled * 3); // Three integers per message
PCumulative.reserve(numProcs + 1); // Similar to Row Pointer vector in CSR data structure
PSizeInfoMessages.reserve(numProcs * 3); // Buffer to hold the Size info message packets
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesMessageBundling: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
PMessageBundle.resize(NumMessagesBundled * 3, -1); // Initialize
PCumulative.resize(numProcs + 1, 0); // Only initialize the counter variable
PSizeInfoMessages.resize(numProcs * 3, 0);
}
PCumulative.reserve(numProcs + 1); // Similar to Row Pointer vector in CSR data structure
PSizeInfoMessages.reserve(numProcs * 3); // Buffer to hold the Size info message packets
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesMessageBundling: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
PMessageBundle.resize(NumMessagesBundled * 3, -1); // Initialize
PCumulative.resize(numProcs + 1, 0); // Only initialize the counter variable
PSizeInfoMessages.resize(numProcs * 3, 0);
}
#pragma omp task depend(inout \
: PCumulative) depend(in \
: PCumulative) depend(in \
: PCounter)
{
for (i = 0; i < numProcs; i++)
PCumulative[i + 1] = PCumulative[i] + PCounter[i];
}
{
for (i = 0; i < numProcs; i++)
PCumulative[i + 1] = PCumulative[i] + PCounter[i];
}
#pragma omp task depend(inout \
: PCounter)
{
// Reuse PCounter to keep track of how many messages were inserted:
for (MilanInt i = 0; i < numProcs; i++) // Changed by Fabio to be an integer, addresses needs to be integers!
PCounter[i] = 0;
}
{
// Reuse PCounter to keep track of how many messages were inserted:
for (MilanInt i = 0; i < numProcs; i++) // Changed by Fabio to be an integer, addresses needs to be integers!
PCounter[i] = 0;
}
// Build the Message Bundle packet:
#pragma omp task depend(in \
: PCounter, QLocalVtx, QGhostVtx, QMsgType, QOwner, PMessageBundle, PCumulative) depend(out \
: myIndex, PMessageBundle, PCounter)
{
for (i = 0; i < NumMessagesBundled; i++)
{
myIndex = (PCumulative[QOwner[i]] + PCounter[QOwner[i]]) * 3;
PMessageBundle[myIndex + 0] = QLocalVtx[i];
PMessageBundle[myIndex + 1] = QGhostVtx[i];
PMessageBundle[myIndex + 2] = QMsgType[i];
PCounter[QOwner[i]]++;
}
}
for (i = 0; i < NumMessagesBundled; i++) {
myIndex = (PCumulative[QOwner[i]] + PCounter[QOwner[i]]) * 3;
PMessageBundle[myIndex + 0] = QLocalVtx[i];
PMessageBundle[myIndex + 1] = QGhostVtx[i];
PMessageBundle[myIndex + 2] = QMsgType[i];
PCounter[QOwner[i]]++;
}
}
// Send the Bundled Messages: Use ISend
#pragma omp task depend(out \
: SRequest, SStatus)
{
try
{
SRequest.reserve(numProcs * 2); // At most two messages per processor
SStatus.reserve(numProcs * 2); // At most two messages per processor
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesLinearSearchImmediateSend: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
}
{
try
{
SRequest.reserve(numProcs * 2); // At most two messages per processor
SStatus.reserve(numProcs * 2); // At most two messages per processor
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesLinearSearchImmediateSend: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
}
// Send the Messages
#pragma omp task depend(inout \
: SRequest, PSizeInfoMessages, PCumulative) depend(out \
: *msgActual, *msgInd)
{
for (i = 0; i < numProcs; i++)
{ // Changed by Fabio to be an integer, addresses needs to be integers!
if (i == myRank) // Do not send anything to yourself
continue;
// Send the Message with information about the size of next message:
// Build the Message Packet:
PSizeInfoMessages[i * 3 + 0] = (PCumulative[i + 1] - PCumulative[i]) * 3; // # of integers in the next message
PSizeInfoMessages[i * 3 + 1] = -1; // Dummy packet
PSizeInfoMessages[i * 3 + 2] = SIZEINFO; // TYPE
// Send a Request (Asynchronous)
for (i = 0; i < numProcs; i++) { // Changed by Fabio to be an integer, addresses needs to be integers!
if (i == myRank) // Do not send anything to yourself
continue;
// Send the Message with information about the size of next message:
// Build the Message Packet:
PSizeInfoMessages[i * 3 + 0] = (PCumulative[i + 1] - PCumulative[i]) * 3; // # of integers in the next message
PSizeInfoMessages[i * 3 + 1] = -1; // Dummy packet
PSizeInfoMessages[i * 3 + 2] = SIZEINFO; // TYPE
// Send a Request (Asynchronous)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending bundled message to process " << i << " size: " << PSizeInfoMessages[i * 3 + 0] << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Sending bundled message to process " << i << " size: " << PSizeInfoMessages[i * 3 + 0] << endl;
fflush(stdout);
#endif
if (PSizeInfoMessages[i * 3 + 0] > 0)
{ // Send only if it is a nonempty packet
MPI_Isend(&PSizeInfoMessages[i * 3 + 0], 3, TypeMap<MilanLongInt>(), i, ComputeTag, comm,
&SRequest[(*msgInd)]);
(*msgActual)++;
(*msgInd)++;
// Now Send the message with the data packet:
if (PSizeInfoMessages[i * 3 + 0] > 0)
{ // Send only if it is a nonempty packet
MPI_Isend(&PSizeInfoMessages[i * 3 + 0], 3, TypeMap<MilanLongInt>(), i, ComputeTag, comm,
&SRequest[(*msgInd)]);
(*msgActual)++;
(*msgInd)++;
// Now Send the message with the data packet:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")SendiFFng Bundle to : " << i << endl;
for (k = (PCumulative[i] * 3); k < (PCumulative[i] * 3 + PSizeInfoMessages[i * 3 + 0]); k++)
cout << PMessageBundle[k] << ",";
cout << endl;
fflush(stdout);
cout << "\n(" << myRank << ")SendiFFng Bundle to : " << i << endl;
for (k = (PCumulative[i] * 3); k < (PCumulative[i] * 3 + PSizeInfoMessages[i * 3 + 0]); k++)
cout << PMessageBundle[k] << ",";
cout << endl;
fflush(stdout);
#endif
MPI_Isend(&PMessageBundle[PCumulative[i] * 3], PSizeInfoMessages[i * 3 + 0],
TypeMap<MilanLongInt>(), i, BundleTag, comm, &SRequest[(*msgInd)]);
(*msgInd)++;
} // End of if size > 0
}
MPI_Isend(&PMessageBundle[PCumulative[i] * 3], PSizeInfoMessages[i * 3 + 0],
TypeMap<MilanLongInt>(), i, BundleTag, comm, &SRequest[(*msgInd)]);
(*msgInd)++;
} // End of if size > 0
}
}
#pragma omp task depend(inout \
@@ -147,64 +146,64 @@ PSizeInfoMessages.resize(numProcs * 3, 0);
{
// Free up temporary memory:
PCumulative.clear();
QLocalVtx.clear();
QGhostVtx.clear();
QMsgType.clear();
QOwner.clear();
PCumulative.clear();
QLocalVtx.clear();
QGhostVtx.clear();
QMsgType.clear();
QOwner.clear();
}
#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;
cout << "\n(" << myRank << ")Number messages already sent in bundles = " << NumMessagesBundled;
if (*numGhostEdges > 0)
cout << "\n(" << myRank << ")Number of Ghost edges = " << *numGhostEdges;
cout << "\n(" << myRank << ")Total number of potential message X 2 = " << *numGhostEdges * 2;
cout << "\n(" << myRank << ")Number messages already sent in bundles = " << NumMessagesBundled;
if (*numGhostEdges > 0)
{
cout << "\n(" << myRank << ")Percentage of total = " << ((double)NumMessagesBundled / (double)(*numGhostEdges * 2)) * 100.0 << "% \n";
cout << "\n(" << myRank << ")Percentage of total = " << ((double)NumMessagesBundled / (double)(*numGhostEdges * 2)) * 100.0 << "% \n";
}
fflush(stdout);
fflush(stdout);
#endif
// Allocate memory for MPI Send messages:
/* WILL COME BACK HERE - NO NEED TO STORE ALL THIS MEMORY !! */
OneMessageSize = 0;
MPI_Pack_size(3, TypeMap<MilanLongInt>(), comm, &OneMessageSize); // Size of one message packet
// How many messages to send?
// Potentially three kinds of messages will be sent/received:
// Request, Success, Failure.
// But only two will be sent from a given processor.
// Substract the number of messages that have already been sent as bundled messages:
numMessagesToSend = (*numGhostEdges) * 2 - NumMessagesBundled;
*BufferSize = (OneMessageSize + MPI_BSEND_OVERHEAD) * numMessagesToSend;
// Allocate memory for MPI Send messages:
/* WILL COME BACK HERE - NO NEED TO STORE ALL THIS MEMORY !! */
OneMessageSize = 0;
MPI_Pack_size(3, TypeMap<MilanLongInt>(), comm, &OneMessageSize); // Size of one message packet
// How many messages to send?
// Potentially three kinds of messages will be sent/received:
// Request, Success, Failure.
// But only two will be sent from a given processor.
// Substract the number of messages that have already been sent as bundled messages:
numMessagesToSend = (*numGhostEdges) * 2 - NumMessagesBundled;
*BufferSize = (OneMessageSize + MPI_BSEND_OVERHEAD) * numMessagesToSend;
}
#pragma omp task depend(out : Buffer) depend(in : *BufferSize)
{
Buffer = 0;
{
Buffer = 0;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Size of One Message from PACK= " << OneMessageSize;
cout << "\n(" << myRank << ")Size of Message overhead = " << MPI_BSEND_OVERHEAD;
cout << "\n(" << myRank << ")Number of Ghost edges = " << *numGhostEdges;
cout << "\n(" << myRank << ")Number of remaining message = " << numMessagesToSend;
cout << "\n(" << myRank << ")BufferSize = " << (*BufferSize);
cout << "\n(" << myRank << ")Attaching Buffer on.. ";
fflush(stdout);
cout << "\n(" << myRank << ")Size of One Message from PACK= " << OneMessageSize;
cout << "\n(" << myRank << ")Size of Message overhead = " << MPI_BSEND_OVERHEAD;
cout << "\n(" << myRank << ")Number of Ghost edges = " << *numGhostEdges;
cout << "\n(" << myRank << ")Number of remaining message = " << numMessagesToSend;
cout << "\n(" << myRank << ")BufferSize = " << (*BufferSize);
cout << "\n(" << myRank << ")Attaching Buffer on.. ";
fflush(stdout);
#endif
if ((*BufferSize) > 0)
{
Buffer = (MilanLongInt *)malloc((*BufferSize)); // Allocate memory
if (Buffer == 0)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesLinearSearch: \n";
cout << "Not enough memory to allocate for send buffer on process " << myRank << "\n";
exit(1);
}
MPI_Buffer_attach(Buffer, *BufferSize); // Attach the Buffer
}
if ((*BufferSize) > 0)
{
Buffer = (MilanLongInt *)malloc((*BufferSize)); // Allocate memory
if (Buffer == 0)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesLinearSearch: \n";
cout << "Not enough memory to allocate for send buffer on process " << myRank << "\n";
exit(1);
}
MPI_Buffer_attach(Buffer, *BufferSize); // Attach the Buffer
}
}
}
}
}
}