You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
amg4psblas/amgprec/impl/aggregator/algoDistEdgeApproxDomEdgesL...

629 lines
28 KiB
C++

#include "MatchBoxPC.h"
// ***********************************************************************
//
// MatchboxP: A C++ library for approximate weighted matching
// Mahantesh Halappanavar (hala@pnnl.gov)
// Pacific Northwest National Laboratory
//
// ***********************************************************************
//
// Copyright (2021) Battelle Memorial Institute
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions
// are met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in the
// documentation and/or other materials provided with the distribution.
//
// 3. Neither the name of the copyright holder nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
// COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
// POSSIBILITY OF SUCH DAMAGE.
//
// ************************************************************************
//////////////////////////////////////////////////////////////////////////////////////
/////////////////////////// DOMINATING EDGES MODEL ///////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////
/* Function : algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMate()
*
* Date : New update: Feb 17, 2019, Richland, Washington.
* Date : Original development: May 17, 2009, E&CS Bldg.
*
* Purpose : Compute Approximate Maximum Weight Matching in Linear Time
*
* Args : inputMatrix - instance of Compressed-Col format of Matrix
* Mate - The Mate array
*
* Returns : By Value: (void)
* By Reference: Mate
*
* Comments : 1/2 Approx Algorithm. Picks the locally available heaviest edge.
* Assumption: The Mate Array is empty.
*/
/*
NLVer = #of vertices, NLEdge = #of edges
CSR/CSC/Compressed format: verLocPtr = Pointer, verLocInd = Index, edgeLocWeight = edge weights (positive real numbers)
verDistance = A vector of size |P|+1 containing the cumulative number of vertices per process
Mate = A vector of size |V_p| (local subgraph) to store the output (matching)
MPI: myRank, numProcs, comm,
Statistics: msgIndSent, msgActualSent, msgPercent : Size: |P| number of processes in the comm-world
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)
*/
#ifdef SERIAL_MPI
#else
// DOUBLE PRECISION VERSION
// WARNING: The vertex block on a given rank is contiguous
void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt *verLocPtr, MilanLongInt *verLocInd,
MilanReal *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *Mate,
MilanInt myRank, MilanInt numProcs, MPI_Comm comm,
MilanLongInt *msgIndSent, MilanLongInt *msgActualSent,
MilanReal *msgPercent,
MilanReal *ph0_time, MilanReal *ph1_time, MilanReal *ph2_time,
MilanLongInt *ph1_card, MilanLongInt *ph2_card)
{
/*
* verDistance: it's a vector long as the number of processors.
* verDistance[i] contains the first node index of the i-th processor
* verDistance[i + 1] contains the last node index of the i-th processor
* NLVer: number of elements in the LocPtr
* NLEdge: number of edges assigned to the current processor
*
* Contains the portion of matrix assigned to the processor in
* Yale notation
* verLocInd: contains the positions on row of the matrix
* verLocPtr: i-th value is the position of the first element on the i-th row and
* i+1-th value is the position of the first element on the i+1-th row
*/
#if !defined(SERIAL_MPI)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Within algoEdgeApproxDominatingEdgesLinearSearchMessageBundling()";
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ") verDistance [" << verDistance[0] << "," << verDistance[1] << "," << verDistance[2] << "," << verDistance[3] << "]";
fflush(stdout);
#endif
#ifdef DEBUG_HANG_
if (myRank == 0)
cout << "\n(" << myRank << ") verDistance [" << verDistance[0] << "," << verDistance[1] << "," << verDistance[2] << "," << verDistance[3] << "]";
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
MPI_Status computeStatus;
MilanLongInt msgActual = 0, msgInd = 0;
MilanReal heaviestEdgeWt = 0.0f; // Assumes positive weight
MilanReal startTime, finishTime;
startTime = MPI_Wtime();
// Data structures for sending and receiving messages:
vector<MilanLongInt> Message; // [ u, v, message_type ]
Message.resize(3, -1);
// Data structures for Message Bundling:
// Although up to two messages can be sent along any cross edge,
// only one message will be sent in the initialization phase -
// one of: REQUEST/FAILURE/SUCCESS
vector<MilanLongInt> QLocalVtx, QGhostVtx, QMsgType;
vector<MilanInt> QOwner; // Changed by Fabio to be an integer, addresses needs to be integers!
MilanLongInt *PCounter = new MilanLongInt[numProcs];
for (int i = 0; i < numProcs; i++)
PCounter[i] = 0;
MilanLongInt NumMessagesBundled = 0;
// TODO when the last computational section will be refactored this could be eliminated
MilanInt ghostOwner = 0; // Changed by Fabio to be an integer, addresses needs to be integers!
MilanLongInt *candidateMate = nullptr;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")NV: " << NLVer << " Edges: " << NLEdge;
fflush(stdout);
cout << "\n(" << myRank << ")StartIndex: " << StartIndex << " EndIndex: " << EndIndex;
fflush(stdout);
#endif
// Other Variables:
MilanLongInt u = -1, v = -1, w = -1, i = 0;
MilanLongInt k = -1, adj1 = -1, adj2 = -1;
MilanLongInt k1 = -1, adj11 = -1, adj12 = -1;
MilanLongInt myCard = 0;
// Build the Ghost Vertex Set: Vg
map<MilanLongInt, MilanLongInt> Ghost2LocalMap; // Map each ghost vertex to a local vertex
vector<MilanLongInt> Counter; // Store the edge count for each ghost vertex
MilanLongInt numGhostVertices = 0, numGhostEdges = 0; // Number of Ghost vertices
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")About to compute Ghost Vertices...";
fflush(stdout);
#endif
#ifdef DEBUG_HANG_
if (myRank == 0)
cout << "\n(" << myRank << ")About to compute Ghost Vertices...";
fflush(stdout);
#endif
// Define Adjacency Lists for Ghost Vertices:
// cout<<"Building Ghost data structures ... \n\n";
vector<MilanLongInt> verGhostPtr, verGhostInd, tempCounter;
// Mate array for ghost vertices:
vector<MilanLongInt> GMate; // Proportional to the number of ghost vertices
MilanLongInt S;
MilanLongInt privateMyCard = 0;
staticQueue U, privateU, privateQLocalVtx, privateQGhostVtx, privateQMsgType, privateQOwner;
vector<MilanLongInt> PCumulative, PMessageBundle, PSizeInfoMessages;
vector<MPI_Request> SRequest; // Requests that are used for each send message
vector<MPI_Status> SStatus; // Status of sent messages, used in MPI_Wait
MilanLongInt MessageIndex = 0; // Pointer for current message
MilanInt BufferSize;
MilanLongInt *Buffer;
// Declare the locks
omp_lock_t MateLock[NLVer];
initialize(NLVer, NLEdge, StartIndex,
EndIndex, &numGhostEdges,
&numGhostVertices, &S,
verLocInd, verLocPtr,
MateLock,
Ghost2LocalMap, Counter,
verGhostPtr, verGhostInd,
tempCounter, GMate,
Message, QLocalVtx,
QGhostVtx, QMsgType, QOwner,
candidateMate, U,
privateU,
privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
finishTime = MPI_Wtime();
*ph0_time = finishTime - startTime; // Time taken for Phase-0: Initialization
startTime = MPI_Wtime();
/////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////// INITIALIZATION /////////////////////////////////////
/////////////////////////////////////////////////////////////////////////////////////////
// Compute the Initial Matching Set:
/*
* OMP PARALLEL_COMPUTE_CANDIDATE_MATE_B has been splitted from
* PARALLEL_PROCESS_EXPOSED_VERTEX_B in order to better parallelize
* the two.
* PARALLEL_COMPUTE_CANDIDATE_MATE_B is now totally parallel.
*/
PARALLEL_COMPUTE_CANDIDATE_MATE_B(NLVer,
verLocPtr,
verLocInd,
myRank,
edgeLocWeight,
candidateMate);
/*
* PARALLEL_PROCESS_EXPOSED_VERTEX_B
* TODO: write comment
*
* TODO: Test when it's actually more efficient to execute this code
* in parallel.
*/
PARALLEL_PROCESS_EXPOSED_VERTEX_B(NLVer,
candidateMate,
verLocInd,
verLocPtr,
StartIndex,
EndIndex,
Mate,
GMate,
Ghost2LocalMap,
edgeLocWeight,
&myCard,
&msgInd,
&NumMessagesBundled,
&S,
verDistance,
PCounter,
Counter,
myRank,
numProcs,
U,
privateU,
QLocalVtx,
QGhostVtx,
QMsgType,
QOwner,
privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
tempCounter.clear(); // Do not need this any more
///////////////////////////////////////////////////////////////////////////////////
/////////////////////////// PROCESS MATCHED VERTICES //////////////////////////////
///////////////////////////////////////////////////////////////////////////////////
processMatchedVertices(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,
MateLock);
/////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////// SEND BUNDLED MESSAGES /////////////////////////////////////
/////////////////////////////////////////////////////////////////////////////////////////
sendBundledMessages(&numGhostEdges,
&BufferSize,
Buffer,
PCumulative,
PMessageBundle,
PSizeInfoMessages,
PCounter,
NumMessagesBundled,
&msgActual,
&MessageIndex,
numProcs,
myRank,
comm,
QLocalVtx,
QGhostVtx,
QMsgType,
QOwner,
SRequest,
SStatus);
///////////////////////// END OF SEND BUNDLED MESSAGES //////////////////////////////////
finishTime = MPI_Wtime();
*ph1_time = finishTime - startTime; // Time taken for Phase-1
*ph1_card = myCard; // Cardinality at the end of Phase-1
startTime = MPI_Wtime();
/////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////// MAIN LOOP //////////////////////////////////////
/////////////////////////////////////////////////////////////////////////////////////////
// Main While Loop:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Entering While(true) loop..";
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
while (true)
{
#ifdef DEBUG_HANG_
if (myRank == 0)
cout << "\n(" << myRank << ") Main loop" << endl;
fflush(stdout);
#endif
///////////////////////////////////////////////////////////////////////////////////
/////////////////////////// PROCESS MATCHED VERTICES //////////////////////////////
///////////////////////////////////////////////////////////////////////////////////
while (/*!Q.empty()*/ !U.empty())
{
// Q.pop_front();
u = U.pop_front(); // Get an element from the queue
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")u: " << u;
fflush(stdout);
#endif
if ((u >= StartIndex) && (u <= EndIndex))
{ // Process Only If a Local Vertex
// Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++)
{
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex))
{ // v is a Local Vertex:
if (Mate[v - StartIndex] >= 0) // v is already matched
continue;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout);
#endif
if (candidateMate[v - StartIndex] == u)
{ // Only if pointing to the matched vertex
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
// Start: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
adj11 = verLocPtr[v - StartIndex];
adj12 = verLocPtr[v - StartIndex + 1];
w = -1;
heaviestEdgeWt = MilanRealMin; // Assign the smallest Value possible first LDBL_MIN
for (k1 = adj11; k1 < adj12; k1++)
{
if ((verLocInd[k1] < StartIndex) || (verLocInd[k1] > EndIndex))
{ // Is it a ghost vertex?
if (GMate[Ghost2LocalMap[verLocInd[k1]]] >= 0) // Already matched
continue;
}
else
{ // A local vertex
if (Mate[verLocInd[k1] - StartIndex] >= 0) // Already matched
continue;
}
if ((edgeLocWeight[k1] > heaviestEdgeWt) ||
((edgeLocWeight[k1] == heaviestEdgeWt) && (w < verLocInd[k1])))
{
heaviestEdgeWt = edgeLocWeight[k1];
w = verLocInd[k1];
}
} // End of for loop
candidateMate[v - StartIndex] = w;
// End: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w;
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)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request 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);
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
// Q.push_back(u);
U.push_back(v);
U.push_back(w);
myCard++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
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++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
{ // no dominating edge found: w == -1
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);
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
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of If (candidateMate[v-StartIndex] == u)
} // End of if ( (v >= StartIndex) && (v <= EndIndex) ) //If Local Vertex:
else
{ // Neighbor v is a ghost vertex
if (candidateMate[NLVer + Ghost2LocalMap[v]] == u)
candidateMate[NLVer + Ghost2LocalMap[v]] = -1;
if (v != Mate[u - StartIndex])
{ // u is a local vertex
// Build the Message Packet:
Message[0] = u; // LOCAL
Message[1] = v; // GHOST
Message[2] = SUCCESS; // TYPE
// Send a Request (Asynchronous)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs);
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
assert(ghostOwner != -1);
assert(ghostOwner != myRank);
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
msgInd++;
msgActual++;
#ifdef DEBUG_GHOST_
if ((u < StartIndex) || (u > EndIndex))
{
cout << "\n(" << myRank << ") " << __LINE__ << " From Send: should not happen: u= " << u << " v= " << v << " StartIndex " << StartIndex << " EndIndex " << EndIndex << endl;
fflush(stdout);
}
#endif
} // End of If( v != Mate[u] )
} // End of Else //A Ghost Vertex
} // End of For Loop adj(u)
} // End of if ( (u >= StartIndex) && (u <= EndIndex) ) //Process Only If a Local Vertex
} // End of while ( /*!Q.empty()*/ !U.empty() )
///////////////////////// END OF PROCESS MATCHED VERTICES /////////////////////////
//// BREAK IF NO MESSAGES EXPECTED /////////
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Deciding whether to break: S= " << S << endl;
#endif
if (S == 0)
{
#ifdef DEBUG_HANG_
cout << "\n(" << myRank << ") Breaking out" << endl;
fflush(stdout);
#endif
break;
}
///////////////////////////////////////////////////////////////////////////////////
/////////////////////////// PROCESS MESSAGES //////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////
processMessages(NLVer,
Mate,
candidateMate,
Ghost2LocalMap,
GMate,
Counter,
StartIndex,
EndIndex,
&myCard,
&msgInd,
&msgActual,
edgeLocWeight,
verDistance,
verLocPtr,
k,
verLocInd,
numProcs,
myRank,
comm,
Message,
numGhostEdges,
u,
v,
&S,
U);
///////////////////////// END OF PROCESS MESSAGES /////////////////////////////////
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Finished Message processing phase: S= " << S;
fflush(stdout);
cout << "\n(" << myRank << ")** SENT : ACTUAL= " << msgActual;
fflush(stdout);
cout << "\n(" << myRank << ")** SENT : INDIVIDUAL= " << msgInd << endl;
fflush(stdout);
#endif
} // End of while (true)
clean(NLVer,
myRank,
MessageIndex,
SRequest,
SStatus,
BufferSize,
Buffer,
msgActual,
msgActualSent,
msgInd,
msgIndSent,
NumMessagesBundled,
msgPercent,
MateLock);
finishTime = MPI_Wtime();
*ph2_time = finishTime - startTime; // Time taken for Phase-2
*ph2_card = myCard; // Cardinality at the end of Phase-2
}
// End of algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMate
#endif
#endif