fdurastante/amg4psblas
1
0
Fork 0
mirror of https://github.com/sfilippone/amg4psblas.git synced 2026-10-06 22:55:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Refactoring, removed all useless Pointer passed in functions

Browse Source
This commit is contained in:
StefanoPetrilli
2022-07-17 13:20:49 -05:00
parent a71fe82752
commit 3e945c75b4
4 changed files with 85 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files
amgprec/impl/aggregator/processCrossEdge.cpp
+2 -2
View File
@@ -1,7 +1,7 @@
#include "MatchBoxPC.h"
void PROCESS_CROSS_EDGE(MilanLongInt *edge,
MilanLongInt *SPtr)
MilanLongInt *S)
{
// Start: PARALLEL_PROCESS_CROSS_EDGE_B
MilanLongInt captureCounter;
@@ -13,7 +13,7 @@ void PROCESS_CROSS_EDGE(MilanLongInt *edge,
if (captureCounter == 0)
#pragma omp atomic
(*SPtr)--; // Decrement S
(*S)--; // Decrement S
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Decrementing S: Ghost vertex " << edge << " has received all its messages";
amgprec/impl/aggregator/processExposedVertex.cpp
+43 -53
View File
@@ -1,45 +1,43 @@
#include "MatchBoxPC.h"
void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
MilanLongInt *candidateMate,
MilanLongInt *verLocInd,
MilanLongInt *verLocPtr,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *Mate,
vector<MilanLongInt> &GMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *NumMessagesBundledPtr,
MilanLongInt *SPtr,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
staticQueue &U,
staticQueue &privateU,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
staticQueue &privateQLocalVtx,
staticQueue &privateQGhostVtx,
staticQueue &privateQMsgType,
staticQueue &privateQOwner)
MilanLongInt *candidateMate,
MilanLongInt *verLocInd,
MilanLongInt *verLocPtr,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *Mate,
vector<MilanLongInt> &GMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *NumMessagesBundled,
MilanLongInt *S,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
staticQueue &U,
staticQueue &privateU,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
staticQueue &privateQLocalVtx,
staticQueue &privateQGhostVtx,
staticQueue &privateQMsgType,
staticQueue &privateQOwner)
{
MilanLongInt v = -1, k = -1, w = -1, adj11 = 0, adj12 = 0, k1 = 0, S = *SPtr;
MilanLongInt myCard = 0, msgInd = 0;
MilanLongInt NumMessagesBundled = 0;
MilanLongInt v = -1, k = -1, w = -1, adj11 = 0, adj12 = 0, k1 = 0;
MilanInt ghostOwner = 0;
#pragma omp parallel private(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(+ \
: msgInd, NumMessagesBundled, myCard, PCounter[:numProcs]) schedule(static)
: PCounter[:numProcs]) schedule(static)
for (v = 0; v < NLVer; v++)
{
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
@@ -76,8 +74,8 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
if (w >= 0)
{
myCard++;
#pragma omp atomic
(*myCard)++;
if ((w < StartIndex) || (w > EndIndex))
{ // w is a ghost vertex
#ifdef PRINT_DEBUG_INFO_
@@ -85,9 +83,10 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
cout << "\n(" << myRank << ")Local is: " << v + StartIndex << " Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
#endif
msgInd++;
NumMessagesBundled++;
#pragma omp atomic
(*msgInd)++;
#pragma omp atomic
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
assert(ghostOwner != -1);
assert(ghostOwner != myRank);
@@ -98,7 +97,6 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
privateQMsgType.push_back(REQUEST);
privateQOwner.push_back(ghostOwner);
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v + StartIndex)
{
@@ -113,9 +111,9 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
fflush(stdout);
#endif
//TODO refactor this!!
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], &S);
// TODO refactor this!!
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], S);
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
@@ -159,9 +157,10 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
fflush(stdout);
#endif
msgInd++;
NumMessagesBundled++;
#pragma omp atomic
(*msgInd)++;
#pragma omp atomic
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
assert(ghostOwner != -1);
assert(ghostOwner != myRank);
@@ -184,14 +183,5 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
privateQMsgType,
privateQOwner);
//TODO move this outside of the parallel region!!
#pragma omp master
{
*myCardPtr = myCard;
*msgIndPtr = msgInd;
*NumMessagesBundledPtr = NumMessagesBundled;
*SPtr = S;
}
} // End of parallel region
}
amgprec/impl/aggregator/processMessages.cpp
+11 -14
View File
@@ -9,9 +9,9 @@ void processMessages(
vector<MilanLongInt> &Counter,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *msgActualPtr,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *msgActual,
MilanReal *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *verLocPtr,
@@ -30,7 +30,7 @@ void processMessages(
MilanInt Sender;
MPI_Status computeStatus;
MilanLongInt bundleSize, myCard = *myCardPtr, msgInd = *msgIndPtr, msgActual = *msgActualPtr, w;
MilanLongInt bundleSize, w;
MilanLongInt adj11, adj12, k1;
MilanLongInt ghostOwner;
int error_codeC;
@@ -188,7 +188,7 @@ void processMessages(
Mate[v - StartIndex] = u; // v is local
U.push_back(v);
U.push_back(u);
myCard++;
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << u << ") " << endl;
fflush(stdout);
@@ -246,15 +246,15 @@ void processMessages(
assert(ghostOwner != myRank);
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
msgInd++;
msgActual++;
(*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++;
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
@@ -272,7 +272,7 @@ void processMessages(
// Q.push_back(u);
U.push_back(v);
U.push_back(w);
myCard++;
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
@@ -303,8 +303,8 @@ void processMessages(
assert(ghostOwner != -1);
assert(ghostOwner != myRank);
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
msgInd++;
msgActual++;
(*msgInd)++;
(*msgActual)++;
} // End of if(GHOST)
} // End of for loop
} // End of Else: w == -1
@@ -324,8 +324,5 @@ void processMessages(
} // End of else: CASE I
}
*myCardPtr = myCard;
*msgIndPtr = msgInd;
*msgActualPtr = msgActual;
return;
}
amgprec/impl/aggregator/sendBundledMessages.cpp
+29 -29
View File
@@ -1,15 +1,15 @@
#include "MatchBoxPC.h"
void sendBundledMessages(MilanLongInt *numGhostEdgesPtr,
MilanInt *BufferSizePtr,
void sendBundledMessages(MilanLongInt *numGhostEdges,
MilanInt *BufferSize,
MilanLongInt *Buffer,
vector<MilanLongInt> &PCumulative,
vector<MilanLongInt> &PMessageBundle,
vector<MilanLongInt> &PSizeInfoMessages,
MilanLongInt *PCounter,
MilanLongInt NumMessagesBundled,
MilanLongInt *msgActualPtr,
MilanLongInt *MessageIndexPtr,
MilanLongInt *msgActual,
MilanLongInt *MessageIndex,
MilanInt numProcs,
MilanInt myRank,
MPI_Comm comm,
@@ -21,8 +21,8 @@ void sendBundledMessages(MilanLongInt *numGhostEdgesPtr,
vector<MPI_Status> &SStatus)
{
MilanLongInt myIndex = 0, msgActual = *msgActualPtr, MessageIndex = *MessageIndexPtr, numGhostEdges = *numGhostEdgesPtr, numMessagesToSend;
MilanInt i = 0, OneMessageSize = 0, BufferSize = *BufferSizePtr;
MilanLongInt myIndex = 0, numMessagesToSend;
MilanInt i = 0, OneMessageSize = 0;
#ifdef DEBUG_HANG_
if (myRank == 0)
@@ -105,7 +105,7 @@ PSizeInfoMessages.resize(numProcs * 3, 0);
// Send the Messages
#pragma omp task depend(inout \
: SRequest, PSizeInfoMessages, PCumulative) depend(out \
: msgActual, MessageIndex)
: *msgActual, *MessageIndex)
{
for (i = 0; i < numProcs; i++)
{ // Changed by Fabio to be an integer, addresses needs to be integers!
@@ -124,9 +124,9 @@ PSizeInfoMessages.resize(numProcs * 3, 0);
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[MessageIndex]);
msgActual++;
MessageIndex++;
&SRequest[(*MessageIndex)]);
(*msgActual)++;
(*MessageIndex)++;
// Now Send the message with the data packet:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")SendiFFng Bundle to : " << i << endl;
@@ -136,8 +136,8 @@ PSizeInfoMessages.resize(numProcs * 3, 0);
fflush(stdout);
#endif
MPI_Isend(&PMessageBundle[PCumulative[i] * 3], PSizeInfoMessages[i * 3 + 0],
TypeMap<MilanLongInt>(), i, BundleTag, comm, &SRequest[MessageIndex]);
MessageIndex++;
TypeMap<MilanLongInt>(), i, BundleTag, comm, &SRequest[(*MessageIndex)]);
(*MessageIndex)++;
} // End of if size > 0
}
}
@@ -154,16 +154,16 @@ PSizeInfoMessages.resize(numProcs * 3, 0);
QOwner.clear();
}
#pragma omp task depend(inout : OneMessageSize, BufferSize) depend(out : numMessagesToSend) depend(in : numGhostEdges)
#pragma omp task depend(inout : OneMessageSize, *BufferSize) depend(out : numMessagesToSend) depend(in : *numGhostEdges)
{
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Number of Ghost edges = " << numGhostEdges;
cout << "\n(" << myRank << ")Total number of potential message X 2 = " << numGhostEdges * 2;
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)
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);
#endif
@@ -177,39 +177,39 @@ PSizeInfoMessages.resize(numProcs * 3, 0);
// 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;
numMessagesToSend = (*numGhostEdges) * 2 - NumMessagesBundled;
*BufferSize = (OneMessageSize + MPI_BSEND_OVERHEAD) * numMessagesToSend;
}
#pragma omp task depend(out : Buffer) depend(in : BufferSize)
#pragma omp task depend(out : Buffer) depend(in : *BufferSize)
{
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 Ghost edges = " << *numGhostEdges;
cout << "\n(" << myRank << ")Number of remaining message = " << numMessagesToSend;
cout << "\n(" << myRank << ")BufferSize = " << BufferSize;
cout << "\n(" << myRank << ")BufferSize = " << (*BufferSize);
cout << "\n(" << myRank << ")Attaching Buffer on.. ";
fflush(stdout);
#endif
if (BufferSize > 0)
if ((*BufferSize) > 0)
{
Buffer = (MilanLongInt *)malloc(BufferSize); // Allocate memory
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
MPI_Buffer_attach(Buffer, *BufferSize); // Attach the Buffer
}
}
}
}
*MessageIndexPtr = MessageIndex;
*msgActualPtr = msgActual;
*numGhostEdgesPtr = numGhostEdges;
*BufferSizePtr = BufferSize;
//*MessageIndexPtr = MessageIndex;
//*msgActualPtr = msgActual;
//*numGhostEdgesPtr = numGhostEdges;
//*BufferSizePtr = BufferSize;
}