Mark* score4K(Dice *dice[])
{
int i, j, count;
Mark *mark = (Mark *)malloc(sizeof(Mark));
/* tests each die value for a 4k or better */
for(i = 1; i <= 6; i++)
{
count = 0;
/* counts the number of occurence of the die */
for(j = 0; j < 5; j++)
if(i == dice[j]->value)
count++;
/* if 4k or better breka */
if(count >= 4)
{
mark->value = sumAll(dice);
break;
}
}
mark->vol = (count >= 4) ? 0 : 1;
return mark;
}
std::string MHDRAPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::PrintLoadBalancingInfo(const Matrix & Ac, const std::string & msgTag) {
std::stringstream ss(std::stringstream::out);
// TODO: provide a option to skip this (to avoid global communication)
// TODO: skip if nproc == 1
//nonzero imbalance
size_t numMyNnz = Ac.getNodeNumEntries();
GO maxNnz, minNnz;
RCP<const Teuchos::Comm<int> > comm = Ac.getRowMap()->getComm();
maxAll(comm,(GO)numMyNnz,maxNnz);
//min nnz over all proc (disallow any processors with 0 nnz)
minAll(comm, (GO)((numMyNnz > 0) ? numMyNnz : maxNnz), minNnz);
double imbalance = ((double) maxNnz) / minNnz;
size_t numMyRows = Ac.getNodeNumRows();
//Check whether Ac is spread over more than one process.
GO numActiveProcesses=0;
sumAll(comm, (GO)((numMyRows > 0) ? 1 : 0), numActiveProcesses);
//min, max, and avg # rows per proc
GO minNumRows, maxNumRows;
double avgNumRows;
maxAll(comm, (GO)numMyRows, maxNumRows);
minAll(comm, (GO)((numMyRows > 0) ? numMyRows : maxNumRows), minNumRows);
assert(numActiveProcesses > 0);
avgNumRows = Ac.getGlobalNumRows() / numActiveProcesses;
ss << msgTag << " # processes with rows = " << numActiveProcesses << std::endl;
ss << msgTag << " min # rows per proc = " << minNumRows << ", max # rows per proc = " << maxNumRows << ", avg # rows per proc = " << avgNumRows << std::endl;
ss << msgTag << " nonzero imbalance = " << imbalance << std::endl;
return ss.str();
}
void LeftoverAggregationAlgorithm<LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::RootCandidates(my_size_t nVertices,
ArrayView<const LO> & vertex2AggId, GraphBase const &graph,
ArrayRCP<LO> &candidates, my_size_t &nCandidates, global_size_t &nCandidatesGlobal) const
{
nCandidates = 0;
for (my_size_t i = 0; i < nVertices; i++ ) {
if (vertex2AggId[i] == MUELU_UNAGGREGATED) {
bool noAggdNeighbors = true;
// neighOfINode is the neighbor node list of node 'iNode'.
ArrayView<const LO> neighOfINode = graph.getNeighborVertices(i);
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int adjacent = *it;
if (vertex2AggId[adjacent] != MUELU_UNAGGREGATED)
noAggdNeighbors = false;
}
if (noAggdNeighbors == true) candidates[nCandidates++] = i;
}
}
sumAll(graph.GetComm(), (GO)nCandidates, nCandidatesGlobal);
} //RootCandidates
void RAPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::CheckMainDiagonal(RCP<Matrix> & Ac) const {
// plausibility check: no zeros on diagonal
RCP<Vector> diagVec = VectorFactory::Build(Ac->getRowMap());
Ac->getLocalDiagCopy(*diagVec);
SC zero = Teuchos::ScalarTraits<SC>::zero(), one = Teuchos::ScalarTraits<SC>::one();
LO lZeroDiags = 0;
Teuchos::ArrayRCP< Scalar > diagVal = diagVec->getDataNonConst(0);
for (size_t r = 0; r < Ac->getRowMap()->getNodeNumElements(); r++) {
if (diagVal[r] == zero) {
lZeroDiags++;
if (repairZeroDiagonals_) {
GO grid = Ac->getRowMap()->getGlobalElement(r);
LO lcid = Ac->getColMap()->getLocalElement(grid);
Teuchos::ArrayRCP<LO> indout(1, lcid);
Teuchos::ArrayRCP<SC> valout(1, one);
Ac->insertLocalValues(r, indout.view(0, indout.size()), valout.view(0, valout.size()));
}
}
}
if (IsPrint(Warnings0)) {
const RCP<const Teuchos::Comm<int> > & comm = Ac->getRowMap()->getComm();
GO lZeroDiagsGO = Teuchos::as<GO>(lZeroDiags); /* LO->GO conversion */
GO gZeroDiags = 0;
sumAll(comm, lZeroDiagsGO, gZeroDiags);
if (repairZeroDiagonals_) GetOStream(Warnings0,0) << "RAPFactory (WARNING): repaired " << gZeroDiags << " zeros on main diagonal of Ac." << std::endl;
else GetOStream(Warnings0,0) << "RAPFactory (WARNING): found " << gZeroDiags << " zeros on main diagonal of Ac." << std::endl;
}
}
// Returns the roll's value for a given scoring choice
const int DiceRoll::rollValue(const ChoiceAction choice) const
{
switch(choice) {
case ACES:
return sumIfEqual(1);
case TWOS:
return sumIfEqual(2);
case THREES:
return sumIfEqual(3);
case FOURS:
return sumIfEqual(4);
case FIVES:
return sumIfEqual(5);
case SIXES:
return sumIfEqual(6);
case THREE_OF_A_KIND:
if (isThreeOfAKind())
return sumAll();
return 0;
case FOUR_OF_A_KIND:
if (isFourOfAKind())
return sumAll();
return 0;
case FULL_HOUSE:
if (isFullHouse())
return 25;
return 0;
case SMALL_STRAIGHT:
if (isSmallStraight())
return 30;
return 0;
case LARGE_STRAIGHT:
if (isLargeStraight())
return 40;
return 0;
case YAHTZEE:
if (isYahtzee())
return 50;
return 0;
case CHANCE:
return sumAll();
case ROLL:
assert(false); // Unscorable!
}
}
Mark* scoreChance(Dice *dice[])
{
Mark *temp = score4K(dice),
*mark = (Mark *)malloc(sizeof(Mark));
mark->value = sumAll(dice);
mark->vol = (temp->value > 0) ? 0 : 1;
free(temp);
return mark;
}
int main(void)
{
int j[] = {4,10,19,42,15,53,36}, capacity = 7;
printf("The smallest value in the vector is: %d\n", minValue(j, capacity));
printf("The product of all the values in the vector is: %d\n", productOfAll(j, capacity));
printf("The maximum value in the vector is: %d\n", maxValue(j, capacity));
printf("The index of the minimum value of the vector is: %d\n", minIndex(j, capacity));
printf("The sum of the vectors is: %d\n", sumAll(j, capacity));
printf("The max value index is: %d\n", maxIndex(j, capacity));
}
GlobalOrdinal Aggregates<LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::GetNumGlobalAggregates() const {
LO nAggregates = GetNumAggregates();
GO nGlobalAggregates; sumAll(vertex2AggId_->getMap()->getComm(), (GO)nAggregates, nGlobalAggregates);
return nGlobalAggregates;
}
void Newton::solveNonLinear() {
casadi_log("Newton::solveNonLinear:begin");
// Set up timers for profiling
double time_zero=0;
double time_start=0;
double time_stop=0;
if (CasadiOptions::profiling && !CasadiOptions::profilingBinary) {
time_zero = getRealTime();
CasadiOptions::profilingLog << "start " << this << ":" <<getOption("name") << std::endl;
}
// Pass the inputs to J
for (int i=0; i<getNumInputs(); ++i) {
if (i!=iin_) jac_.setInput(input(i), i);
}
// Aliases
DMatrix &u = output(iout_);
DMatrix &J = jac_.output(0);
DMatrix &F = jac_.output(1+iout_);
// Perform the Newton iterations
int iter=0;
bool success = true;
while (true) {
// Break if maximum number of iterations already reached
if (iter >= max_iter_) {
log("evaluate", "Max. iterations reached.");
stats_["return_status"] = "max_iteration_reached";
success = false;
break;
}
// Start a new iteration
iter++;
// Print progress
if (monitored("step") || monitored("stepsize")) {
std::cout << "Step " << iter << "." << std::endl;
}
if (monitored("step")) {
std::cout << " u = " << u << std::endl;
}
// Use u to evaluate J
jac_.setInput(u, iin_);
for (int i=0; i<getNumInputs(); ++i)
if (i!=iin_) jac_.setInput(input(i), i);
if (CasadiOptions::profiling) {
time_start = getRealTime(); // Start timer
}
jac_.evaluate();
// Write out profiling information
if (CasadiOptions::profiling && !CasadiOptions::profilingBinary) {
time_stop = getRealTime(); // Stop timer
CasadiOptions::profilingLog
<< (time_stop-time_start)*1e6 << " ns | "
<< (time_stop-time_zero)*1e3 << " ms | "
<< this << ":" << getOption("name") << ":0|" << jac_.get() << ":"
<< jac_.getOption("name") << "|evaluate jacobian" << std::endl;
}
if (monitored("F")) std::cout << " F = " << F << std::endl;
if (monitored("normF"))
std::cout << " F (min, max, 1-norm, 2-norm) = "
<< (*std::min_element(F.data().begin(), F.data().end()))
<< ", " << (*std::max_element(F.data().begin(), F.data().end()))
<< ", " << sumAll(fabs(F)) << ", " << sqrt(sumAll(F*F)) << std::endl;
if (monitored("J")) std::cout << " J = " << J << std::endl;
double abstol = 0;
if (numeric_limits<double>::infinity() != abstol_) {
abstol = std::max((*std::max_element(F.data().begin(),
F.data().end())),
-(*std::min_element(F.data().begin(),
F.data().end())));
if (abstol <= abstol_) {
casadi_log("Converged to acceptable tolerance - abstol: " << abstol_);
break;
}
}
// Prepare the linear solver with J
linsol_.setInput(J, LINSOL_A);
if (CasadiOptions::profiling) {
time_start = getRealTime(); // Start timer
}
linsol_.prepare();
// Write out profiling information
if (CasadiOptions::profiling && !CasadiOptions::profilingBinary) {
time_stop = getRealTime(); // Stop timer
CasadiOptions::profilingLog
<< (time_stop-time_start)*1e6 << " ns | "
<< (time_stop-time_zero)*1e3 << " ms | "
<< this << ":" << getOption("name")
<< ":1||prepare linear system" << std::endl;
}
if (CasadiOptions::profiling) {
time_start = getRealTime(); // Start timer
}
// Solve against F
linsol_.solve(&F.front(), 1, false);
if (CasadiOptions::profiling && !CasadiOptions::profilingBinary) {
time_stop = getRealTime(); // Stop timer
CasadiOptions::profilingLog
<< (time_stop-time_start)*1e6 << " ns | "
<< (time_stop-time_zero)*1e3 << " ms | "
<< this << ":" << getOption("name") << ":2||solve linear system" << std::endl;
}
if (monitored("step")) {
std::cout << " step = " << F << std::endl;
}
double abstolStep=0;
if (numeric_limits<double>::infinity() != abstolStep_) {
abstolStep = std::max((*std::max_element(F.data().begin(),
F.data().end())),
-(*std::min_element(F.data().begin(),
F.data().end())));
if (monitored("stepsize")) {
std::cout << " stepsize = " << abstolStep << std::endl;
}
if (abstolStep <= abstolStep_) {
casadi_log("Converged to acceptable tolerance - abstolStep: " << abstolStep_);
break;
}
}
if (print_iteration_) {
// Only print iteration header once in a while
if (iter % 10==0) {
printIteration(std::cout);
}
// Print iteration information
printIteration(std::cout, iter, abstol, abstolStep);
}
// Update Xk+1 = Xk - J^(-1) F
std::transform(u.begin(), u.end(), F.begin(), u.begin(), std::minus<double>());
// Get auxiliary outputs
for (int i=0; i<getNumOutputs(); ++i) {
if (i!=iout_) jac_.getOutput(output(i), 1+i);
}
}
// Store the iteration count
if (gather_stats_) stats_["iter"] = iter;
if (success) stats_["return_status"] = "success";
// Factorization up-to-date
fact_up_to_date_ = true;
casadi_log("Newton::solveNonLinear():end after " << iter << " steps");
}
void UncoupledAggregationFactory<LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Build(Level ¤tLevel) const {
FactoryMonitor m(*this, "Build", currentLevel);
ParameterList pL = GetParameterList();
bDefinitionPhase_ = false; // definition phase is finished, now all aggregation algorithm information is fixed
// define aggregation algorithms
RCP<const FactoryBase> graphFact = GetFactory("Graph");
// TODO Can we keep different aggregation algorithms over more Build calls?
algos_.clear();
if (pL.get<std::string>("aggregation: mode") == "old") {
if (pL.get<bool>("UseOnePtAggregationAlgorithm") == true) algos_.push_back(rcp(new OnePtAggregationAlgorithm (graphFact)));
if (pL.get<bool>("UsePreserveDirichletAggregationAlgorithm") == true) algos_.push_back(rcp(new PreserveDirichletAggregationAlgorithm (graphFact)));
if (pL.get<bool>("UseUncoupledAggregationAlgorithm") == true) algos_.push_back(rcp(new AggregationPhase1Algorithm (graphFact)));
if (pL.get<bool>("UseMaxLinkAggregationAlgorithm") == true) algos_.push_back(rcp(new MaxLinkAggregationAlgorithm (graphFact)));
if (pL.get<bool>("UseEmergencyAggregationAlgorithm") == true) algos_.push_back(rcp(new EmergencyAggregationAlgorithm (graphFact)));
algos_.push_back(rcp(new IsolatedNodeAggregationAlgorithm (graphFact)));
} else {
if (pL.get<bool>("aggregation: preserve Dirichlet points") == true) algos_.push_back(rcp(new PreserveDirichletAggregationAlgorithm (graphFact)));
if (pL.get<bool>("aggregation: enable phase 1" ) == true) algos_.push_back(rcp(new AggregationPhase1Algorithm (graphFact)));
if (pL.get<bool>("aggregation: enable phase 2a") == true) algos_.push_back(rcp(new AggregationPhase2aAlgorithm (graphFact)));
if (pL.get<bool>("aggregation: enable phase 2b") == true) algos_.push_back(rcp(new AggregationPhase2bAlgorithm (graphFact)));
if (pL.get<bool>("aggregation: enable phase 3" ) == true) algos_.push_back(rcp(new AggregationPhase3Algorithm (graphFact)));
algos_.push_back(rcp(new IsolatedNodeAggregationAlgorithm (graphFact)));
}
std::string mapOnePtName = pL.get<std::string>("OnePt aggregate map name");
RCP<const Map> OnePtMap;
if (mapOnePtName.length()) {
RCP<const FactoryBase> mapOnePtFact = GetFactory("OnePt aggregate map factory");
OnePtMap = currentLevel.Get<RCP<const Map> >(mapOnePtName, mapOnePtFact.get());
}
RCP<const GraphBase> graph = Get< RCP<GraphBase> >(currentLevel, "Graph");
// Build
RCP<Aggregates> aggregates = rcp(new Aggregates(*graph));
aggregates->setObjectLabel("UC");
const LO numRows = graph->GetNodeNumVertices();
// construct aggStat information
std::vector<unsigned> aggStat(numRows, READY);
ArrayRCP<const bool> dirichletBoundaryMap = graph->GetBoundaryNodeMap();
if (dirichletBoundaryMap != Teuchos::null)
for (LO i = 0; i < numRows; i++)
if (dirichletBoundaryMap[i] == true)
aggStat[i] = BOUNDARY;
LO nDofsPerNode = Get<LO>(currentLevel, "DofsPerNode");
GO indexBase = graph->GetDomainMap()->getIndexBase();
if (OnePtMap != Teuchos::null) {
for (LO i = 0; i < numRows; i++) {
// reconstruct global row id (FIXME only works for contiguous maps)
GO grid = (graph->GetDomainMap()->getGlobalElement(i)-indexBase) * nDofsPerNode + indexBase;
for (LO kr = 0; kr < nDofsPerNode; kr++)
if (OnePtMap->isNodeGlobalElement(grid + kr))
aggStat[i] = ONEPT;
}
}
const RCP<const Teuchos::Comm<int> > comm = graph->GetComm();
GO numGlobalRows = 0;
if (IsPrint(Statistics1))
sumAll(comm, as<GO>(numRows), numGlobalRows);
LO numNonAggregatedNodes = numRows;
GO numGlobalAggregatedPrev = 0, numGlobalAggsPrev = 0;
for (size_t a = 0; a < algos_.size(); a++) {
std::string phase = algos_[a]->description();
SubFactoryMonitor sfm(*this, "Algo \"" + phase + "\"", currentLevel);
algos_[a]->BuildAggregates(pL, *graph, *aggregates, aggStat, numNonAggregatedNodes);
if (IsPrint(Statistics1)) {
GO numLocalAggregated = numRows - numNonAggregatedNodes, numGlobalAggregated = 0;
GO numLocalAggs = aggregates->GetNumAggregates(), numGlobalAggs = 0;
sumAll(comm, numLocalAggregated, numGlobalAggregated);
sumAll(comm, numLocalAggs, numGlobalAggs);
double aggPercent = 100*as<double>(numGlobalAggregated)/as<double>(numGlobalRows);
if (aggPercent > 99.99 && aggPercent < 100.00) {
// Due to round off (for instance, for 140465733/140466897), we could
// get 100.00% display even if there are some remaining nodes. This
// is bad from the users point of view. It is much better to change
// it to display 99.99%.
aggPercent = 99.99;
}
GetOStream(Statistics1) << " aggregated : " << (numGlobalAggregated - numGlobalAggregatedPrev) << " (phase), " << std::fixed
<< std::setprecision(2) << numGlobalAggregated << "/" << numGlobalRows << " [" << aggPercent << "%] (total)\n"
<< " remaining : " << numGlobalRows - numGlobalAggregated << "\n"
<< " aggregates : " << numGlobalAggs-numGlobalAggsPrev << " (phase), " << numGlobalAggs << " (total)" << std::endl;
numGlobalAggregatedPrev = numGlobalAggregated;
numGlobalAggsPrev = numGlobalAggs;
}
}
TEUCHOS_TEST_FOR_EXCEPTION(numNonAggregatedNodes, Exceptions::RuntimeError, "MueLu::UncoupledAggregationFactory::Build: Leftover nodes found! Error!");
aggregates->AggregatesCrossProcessors(false);
Set(currentLevel, "Aggregates", aggregates);
GetOStream(Statistics0) << aggregates->description() << std::endl;
}
void LocalAggregationAlgorithm<LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::CoarsenUncoupled(GraphBase const & graph, Aggregates & aggregates) const {
Monitor m(*this, "Coarsen Uncoupled");
std::string orderingType;
switch(ordering_) {
case NATURAL:
orderingType="Natural";
break;
case RANDOM:
orderingType="Random";
break;
case GRAPH:
orderingType="Graph";
break;
default:
break;
}
GetOStream(Runtime1) << "Ordering: " << orderingType << std::endl;
GetOStream(Runtime1) << "Min nodes per aggregate: " << minNodesPerAggregate_ << std::endl;
GetOStream(Runtime1) << "Max nbrs already selected: " << maxNeighAlreadySelected_ << std::endl;
/* Create Aggregation object */
my_size_t nAggregates = 0;
/* ============================================================= */
/* aggStat indicates whether this node has been aggreated, and */
/* vertex2AggId stores the aggregate number where this node has */
/* been aggregated into. */
/* ============================================================= */
Teuchos::ArrayRCP<NodeState> aggStat;
const my_size_t nRows = graph.GetNodeNumVertices();
if (nRows > 0) aggStat = Teuchos::arcp<NodeState>(nRows);
for ( my_size_t i = 0; i < nRows; ++i ) aggStat[i] = READY;
/* ============================================================= */
/* Phase 1 : */
/* for all nodes, form a new aggregate with its neighbors */
/* if the number of its neighbors having been aggregated does */
/* not exceed a given threshold */
/* (GetMaxNeighAlreadySelected() = 0 ===> Vanek's scheme) */
/* ============================================================= */
/* some general variable declarations */
Teuchos::ArrayRCP<LO> randomVector;
RCP<MueLu::LinkedList> nodeList; /* list storing the next node to pick as a root point for ordering_ == GRAPH */
MueLu_SuperNode *aggHead=NULL, *aggCurrent=NULL, *supernode=NULL;
/**/
if ( ordering_ == RANDOM ) /* random ordering */
{
//TODO: could be stored in a class that respect interface of LinkedList
randomVector = Teuchos::arcp<LO>(nRows); //size_t or int ?-> to be propagated
for (my_size_t i = 0; i < nRows; ++i) randomVector[i] = i;
RandomReorder(randomVector);
}
else if ( ordering_ == GRAPH ) /* graph ordering */
{
nodeList = rcp(new MueLu::LinkedList());
nodeList->Add(0);
}
/* main loop */
{
LO iNode = 0;
LO iNode2 = 0;
Teuchos::ArrayRCP<LO> vertex2AggId = aggregates.GetVertex2AggId()->getDataNonConst(0); // output only: contents ignored
while (iNode2 < nRows)
{
/*------------------------------------------------------ */
/* pick the next node to aggregate */
/*------------------------------------------------------ */
if ( ordering_ == NATURAL ) iNode = iNode2++;
else if ( ordering_ == RANDOM ) iNode = randomVector[iNode2++];
else if ( ordering_ == GRAPH )
{
if ( nodeList->IsEmpty() )
{
for ( int jNode = 0; jNode < nRows; ++jNode )
{
if ( aggStat[jNode] == READY )
{
nodeList->Add(jNode); //TODO optim: not necessary to create a node. Can just set iNode value and skip the end
break;
}
}
}
if ( nodeList->IsEmpty() ) break; /* end of the while loop */ //TODO: coding style :(
iNode = nodeList->Pop();
}
else {
throw(Exceptions::RuntimeError("CoarsenUncoupled: bad aggregation ordering option"));
}
/*------------------------------------------------------ */
/* consider further only if the node is in READY mode */
/*------------------------------------------------------ */
if ( aggStat[iNode] == READY )
{
// neighOfINode is the neighbor node list of node 'iNode'.
Teuchos::ArrayView<const LO> neighOfINode = graph.getNeighborVertices(iNode);
typename Teuchos::ArrayView<const LO>::size_type length = neighOfINode.size();
supernode = new MueLu_SuperNode;
try {
supernode->list = Teuchos::arcp<int>(length+1);
} catch (std::bad_alloc&) {
TEUCHOS_TEST_FOR_EXCEPTION(true, Exceptions::RuntimeError, "MueLu::LocalAggregationAlgorithm::CoarsenUncoupled(): Error: couldn't allocate memory for supernode! length=" + Teuchos::toString(length));
}
supernode->maxLength = length;
supernode->length = 1;
supernode->list[0] = iNode;
int selectFlag = 1;
{
/*--------------------------------------------------- */
/* count the no. of neighbors having been aggregated */
/*--------------------------------------------------- */
int count = 0;
for (typename Teuchos::ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it)
{
int index = *it;
if ( index < nRows )
{
if ( aggStat[index] == READY ||
aggStat[index] == NOTSEL )
supernode->list[supernode->length++] = index;
else count++;
}
}
/*--------------------------------------------------- */
/* if there are too many neighbors aggregated or the */
/* number of nodes in the new aggregate is too few, */
/* don't do this one */
/*--------------------------------------------------- */
if ( count > GetMaxNeighAlreadySelected() ) selectFlag = 0;
}
// Note: the supernode length is actually 1 more than the
// number of nodes in the candidate aggregate. The
// root is counted twice. I'm not sure if this is
// a bug or a feature ... so I'll leave it and change
// < to <= in the if just below.
if (selectFlag != 1 ||
supernode->length <= GetMinNodesPerAggregate())
{
aggStat[iNode] = NOTSEL;
delete supernode;
if ( ordering_ == GRAPH ) /* if graph ordering */
{
for (typename Teuchos::ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it)
{
int index = *it;
if ( index < nRows && aggStat[index] == READY )
{
nodeList->Add(index);
}
}
}
}
else
{
aggregates.SetIsRoot(iNode);
for ( int j = 0; j < supernode->length; ++j )
{
int jNode = supernode->list[j];
aggStat[jNode] = SELECTED;
vertex2AggId[jNode] = nAggregates;
if ( ordering_ == GRAPH ) /* if graph ordering */
{
Teuchos::ArrayView<const LO> neighOfJNode = graph.getNeighborVertices(jNode);
for (typename Teuchos::ArrayView<const LO>::const_iterator it = neighOfJNode.begin(); it != neighOfJNode.end(); ++it)
{
int index = *it;
if ( index < nRows && aggStat[index] == READY )
{
nodeList->Add(index);
}
}
}
}
supernode->next = NULL;
supernode->index = nAggregates;
if ( nAggregates == 0 )
{
aggHead = supernode;
aggCurrent = supernode;
}
else
{
aggCurrent->next = supernode;
aggCurrent = supernode;
}
nAggregates++;
// unused aggCntArray[nAggregates] = supernode->length;
}
}
} // end of 'for'
// views on distributed vectors are freed here.
} // end of 'main loop'
nodeList = Teuchos::null;
/* Update aggregate object */
aggregates.SetNumAggregates(nAggregates);
/* Verbose */
{
const RCP<const Teuchos::Comm<int> > & comm = graph.GetComm();
if (IsPrint(Warnings0)) {
GO localReady=0, globalReady;
// Compute 'localReady'
for ( my_size_t i = 0; i < nRows; ++i )
if (aggStat[i] == READY) localReady++;
// Compute 'globalReady'
sumAll(comm, localReady, globalReady);
if(globalReady > 0)
GetOStream(Warnings0) << "Warning: " << globalReady << " READY nodes left" << std::endl;
}
if (IsPrint(Statistics1)) {
// Compute 'localSelected'
LO localSelected=0;
for ( my_size_t i = 0; i < nRows; ++i )
if ( aggStat[i] == SELECTED ) localSelected++;
// Compute 'globalSelected'
GO globalSelected; sumAll(comm, (GO)localSelected, globalSelected);
// Compute 'globalNRows'
GO globalNRows; sumAll(comm, (GO)nRows, globalNRows);
GetOStream(Statistics1) << "Nodes aggregated = " << globalSelected << " (" << globalNRows << ")" << std::endl;
}
if (IsPrint(Statistics1)) {
GO nAggregatesGlobal; sumAll(comm, (GO)nAggregates, nAggregatesGlobal);
GetOStream(Statistics1) << "Total aggregates = " << nAggregatesGlobal << std::endl;
}
} // verbose
/* ------------------------------------------------------------- */
/* clean up */
/* ------------------------------------------------------------- */
aggCurrent = aggHead;
while ( aggCurrent != NULL )
{
supernode = aggCurrent;
aggCurrent = aggCurrent->next;
delete supernode;
}
} // CoarsenUncoupled
void UncoupledAggregationFactory<LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Build(Level ¤tLevel) const {
FactoryMonitor m(*this, "Build", currentLevel);
const ParameterList& pL = GetParameterList();
bDefinitionPhase_ = false; // definition phase is finished, now all aggregation algorithm information is fixed
bool bUseOnePtAggregationAlgorithm = pL.get<bool>("UseOnePtAggregationAlgorithm");
bool bUseSmallAggregationAlgorithm = pL.get<bool>("UseSmallAggregatesAggregationAlgorithm");
bool bUsePreserveDirichletAggregationAlgorithm = pL.get<bool>("UsePreserveDirichletAggregationAlgorithm");
bool bUseUncoupledAggregationAglorithm = pL.get<bool>("UseUncoupledAggregationAlgorithm");
bool bUseMaxLinkAggregationAlgorithm = pL.get<bool>("UseMaxLinkAggregationAlgorithm");
bool bUseIsolatedNodeAggregationAglorithm = pL.get<bool>("UseIsolatedNodeAggregationAlgorithm");
bool bUseEmergencyAggregationAlgorithm = pL.get<bool>("UseEmergencyAggregationAlgorithm");
// define aggregation algorithms
RCP<const FactoryBase> graphFact = GetFactory("Graph");
// TODO Can we keep different aggregation algorithms over more Build calls?
algos_.clear();
if (bUseOnePtAggregationAlgorithm) algos_.push_back(rcp(new OnePtAggregationAlgorithm (graphFact)));
if (bUseSmallAggregationAlgorithm) algos_.push_back(rcp(new SmallAggregationAlgorithm (graphFact)));
if (bUseUncoupledAggregationAglorithm) algos_.push_back(rcp(new UncoupledAggregationAlgorithm (graphFact)));
if (bUseMaxLinkAggregationAlgorithm) algos_.push_back(rcp(new MaxLinkAggregationAlgorithm (graphFact)));
if (bUsePreserveDirichletAggregationAlgorithm) algos_.push_back(rcp(new PreserveDirichletAggregationAlgorithm (graphFact)));
if (bUseIsolatedNodeAggregationAglorithm) algos_.push_back(rcp(new IsolatedNodeAggregationAlgorithm (graphFact)));
if (bUseEmergencyAggregationAlgorithm) algos_.push_back(rcp(new EmergencyAggregationAlgorithm (graphFact)));
std::string mapOnePtName = pL.get<std::string>("OnePt aggregate map name"), mapSmallAggName = pL.get<std::string>("SmallAgg aggregate map name");
RCP<const Map> OnePtMap, SmallAggMap;
if (mapOnePtName.length()) {
RCP<const FactoryBase> mapOnePtFact = GetFactory("OnePt aggregate map factory");
OnePtMap = currentLevel.Get<RCP<const Map> >(mapOnePtName, mapOnePtFact.get());
}
if (mapSmallAggName.length()) {
RCP<const FactoryBase> mapSmallAggFact = GetFactory("SmallAgg aggregate map factory");
SmallAggMap = currentLevel.Get<RCP<const Map> >(mapSmallAggName, mapSmallAggFact.get());
}
RCP<const GraphBase> graph = Get< RCP<GraphBase> >(currentLevel, "Graph");
// Build
RCP<Aggregates> aggregates = rcp(new Aggregates(*graph));
aggregates->setObjectLabel("UC");
const LO nRows = graph->GetNodeNumVertices();
// construct aggStat information
std::vector<unsigned> aggStat(nRows, NodeStats::READY);
ArrayRCP<const bool> dirichletBoundaryMap = graph->GetBoundaryNodeMap();
if (dirichletBoundaryMap != Teuchos::null) {
for (LO i = 0; i < nRows; i++)
if (dirichletBoundaryMap[i] == true)
aggStat[i] = NodeStats::BOUNDARY;
}
LO nDofsPerNode = Get<LO>(currentLevel, "DofsPerNode");
GO indexBase = graph->GetDomainMap()->getIndexBase();
if (SmallAggMap != Teuchos::null || OnePtMap != Teuchos::null) {
for (LO i = 0; i < nRows; i++) {
// reconstruct global row id (FIXME only works for contiguous maps)
GO grid = (graph->GetDomainMap()->getGlobalElement(i)-indexBase) * nDofsPerNode + indexBase;
if (SmallAggMap != null) {
for (LO kr = 0; kr < nDofsPerNode; kr++) {
if (SmallAggMap->isNodeGlobalElement(grid + kr))
aggStat[i] = MueLu::NodeStats::SMALLAGG;
}
}
if (OnePtMap != null) {
for (LO kr = 0; kr < nDofsPerNode; kr++) {
if (OnePtMap->isNodeGlobalElement(grid + kr))
aggStat[i] = MueLu::NodeStats::ONEPT;
}
}
}
}
const RCP<const Teuchos::Comm<int> > comm = graph->GetComm();
GO numGlobalRows = 0;
if (IsPrint(Statistics1))
sumAll(comm, as<GO>(nRows), numGlobalRows);
LO numNonAggregatedNodes = nRows;
GO numGlobalAggregatedPrev = 0, numGlobalAggsPrev = 0;
for (size_t a = 0; a < algos_.size(); a++) {
std::string phase = algos_[a]->description();
SubFactoryMonitor sfm(*this, "Algo \"" + phase + "\"", currentLevel);
algos_[a]->BuildAggregates(pL, *graph, *aggregates, aggStat, numNonAggregatedNodes);
if (IsPrint(Statistics1)) {
GO numLocalAggregated = nRows - numNonAggregatedNodes, numGlobalAggregated = 0;
GO numLocalAggs = aggregates->GetNumAggregates(), numGlobalAggs = 0;
sumAll(comm, numLocalAggregated, numGlobalAggregated);
sumAll(comm, numLocalAggs, numGlobalAggs);
double aggPercent = 100*as<double>(numGlobalAggregated)/as<double>(numGlobalRows);
GetOStream(Statistics1) << " aggregated : " << (numGlobalAggregated - numGlobalAggregatedPrev) << " (phase), " << std::fixed
<< std::setprecision(2) << numGlobalAggregated << "/" << numGlobalRows << " [" << aggPercent << "%] (total)\n"
<< " remaining : " << numGlobalRows - numGlobalAggregated << "\n"
<< " aggregates : " << numGlobalAggs-numGlobalAggsPrev << " (phase), " << numGlobalAggs << " (total)" << std::endl;
numGlobalAggregatedPrev = numGlobalAggregated;
numGlobalAggsPrev = numGlobalAggs;
}
}
TEUCHOS_TEST_FOR_EXCEPTION(numNonAggregatedNodes, Exceptions::RuntimeError, "MueLu::UncoupledAggregationFactory::Build: Leftover nodes found! Error!");
aggregates->AggregatesCrossProcessors(false);
Set(currentLevel, "Aggregates", aggregates);
GetOStream(Statistics0) << aggregates->description() << std::endl;
}
int main(int argc, char **argv) {
int numbers[5] = {3, 4, 1, 7, 4};
int sum = sumAll(numbers, 5);
printf("sum is: %d\n", sum);
return 0;
}
Int_t main(Int_t argc, Char_t *argv[])
{
ROOT::Mpi::TEnvironment env(argc, argv);
ROOT::Mpi::TIntraCommunicator world;
TVectorT<Double_t> mResult;
Double_t fScalarResult;
TVectorT<Double_t> v1(elements);
TVectorT<Double_t> v2(elements);
for (Int_t i = 0; i < elements; i++) {
v1[i] = i + (i + world.Size());
v2[i] = i * (i + world.Size());
}
///////////////////////////////////////////////
//Testing methdos with results in single Rank//
///////////////////////////////////////////////
ROOT::Mpi::Math::TVectorTWrapper<Double_t> add(v1);
add.Addition(v2, root);
ROOT::Mpi::Math::TVectorTWrapper<Double_t> sub(v1);
sub.Subtraction(v2, root);
ROOT::Mpi::Math::TVectorTWrapper<Double_t> dot(v1);
dot.Dot(v2, root);
ROOT::Mpi::Math::TVectorTWrapper<Double_t> norm2Sqr(v1);
norm2Sqr.Norm2Sqr(root);
ROOT::Mpi::Math::TVectorTWrapper<Double_t> norm1(v1);
norm1.Norm1(root);
ROOT::Mpi::Math::TVectorTWrapper<Double_t> min(v1);
min.Min(root);
ROOT::Mpi::Math::TVectorTWrapper<Double_t> max(v1);
max.Max(root);
ROOT::Mpi::Math::TVectorTWrapper<Double_t> normalize(v1);
normalize.Normalize(root);
ROOT::Mpi::Math::TVectorTWrapper<Double_t> sum(v1);
sum.Sum(root);
if (world.Rank() == root) {
add.GetResult(mResult);
MpiCompareTVectorTest(mResult, v1 + v2, world.Rank(), "Vector Addition Single");
sub.GetResult(mResult);
MpiCompareTVectorTest(mResult, v1 - v2, world.Rank(), "Vector Subtraction Single");
dot.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, Dot(v1, v2) , world.Rank(), "Vector Dot Product Single");
norm2Sqr.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v1.Norm2Sqr() , world.Rank(), "Vector Norm2Sqr Single");
norm1.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v1.Norm1() , world.Rank(), "Vector Norm1 Single");
min.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v1.Min() , world.Rank(), "Vector Min Single");
max.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v1.Max() , world.Rank(), "Vector Max Single");
normalize.GetResult(mResult);
MpiCompareTest(mResult.Norm2Sqr(), ((1 / TMath::Sqrt(v1.Norm2Sqr()))*v1).Norm2Sqr() , world.Rank(), "Vector Normalize Single");
sum.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v1.Sum(), world.Rank(), "Vector Sum Single");
}
///////////////////////////////////////////////
//Testing methdos with results in all ranks //
///////////////////////////////////////////////
ROOT::Mpi::Math::TVectorTWrapper<Double_t> addAll(v1);
add.Addition(v2);
ROOT::Mpi::Math::TVectorTWrapper<Double_t> subAll(v1);
sub.Subtraction(v2);
add.GetResult(mResult);
MpiCompareTVectorTest(mResult, v1 + v2, world.Rank(), "Vector Addition All");
sub.GetResult(mResult);
MpiCompareTVectorTest(mResult, v1 - v2, world.Rank(), "Vector Subtraction All");
ROOT::Mpi::Math::TVectorTWrapper<Double_t> dotAll(v1);
dotAll.Dot(v2);
dotAll.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, Dot(v1, v2) , world.Rank(), "Vector Dot Product All");
ROOT::Mpi::Math::TVectorTWrapper<Double_t> norm2SqrAll(v2);
norm2SqrAll.Norm2Sqr();
norm2SqrAll.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v2.Norm2Sqr() , world.Rank(), "Vector Norm2Sqr All");
ROOT::Mpi::Math::TVectorTWrapper<Double_t> norm1All(v2);
norm1All.Norm1();
norm1All.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v2.Norm1() , world.Rank(), "Vector Norm1 All");
ROOT::Mpi::Math::TVectorTWrapper<Double_t> minAll(v2);
minAll.Min();
minAll.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v2.Min() , world.Rank(), "Vector Min All");
ROOT::Mpi::Math::TVectorTWrapper<Double_t> maxAll(v2);
maxAll.Max();
maxAll.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v2.Max() , world.Rank(), "Vector Max All");
ROOT::Mpi::Math::TVectorTWrapper<Double_t> normalizeAll(v2);
normalizeAll.Normalize();
normalizeAll.GetResult(mResult);
//show if the vector is normalize, then Norm2Sqr of result is near to 1
MpiCompareTest(mResult.Norm2Sqr(), ((1 / TMath::Sqrt(v2.Norm2Sqr()))*v2).Norm2Sqr() , world.Rank(), "Vector Normalize All");
ROOT::Mpi::Math::TVectorTWrapper<Double_t> sumAll(v2);
sumAll.Sum();
sumAll.GetResult(fScalarResult);
MpiCompareTest(fScalarResult, v2.Sum() , world.Rank(), "Vector Sum All");
return 0;
}
void LeftoverAggregationAlgorithm<LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::AggregateLeftovers(GraphBase const &graph, Aggregates &aggregates) const {
Monitor m(*this, "AggregateLeftovers");
my_size_t nVertices = graph.GetNodeNumVertices();
int exp_nRows = aggregates.GetMap()->getNodeNumElements(); // Tentative fix... was previously exp_nRows = nVertices + graph.GetNodeNumGhost();
int myPid = graph.GetComm()->getRank();
my_size_t nAggregates = aggregates.GetNumAggregates();
int minNodesPerAggregate = GetMinNodesPerAggregate();
const RCP<const Map> nonUniqueMap = aggregates.GetMap(); //column map of underlying graph
const RCP<const Map> uniqueMap = graph.GetDomainMap();
MueLu::CoupledAggregationCommHelper<LO,GO,NO,LMO> myWidget(uniqueMap, nonUniqueMap);
//TODO JJH We want to skip this call
RCP<Xpetra::Vector<double,LO,GO,NO> > distWeights = Xpetra::VectorFactory<double,LO,GO,NO>::Build(nonUniqueMap);
// Aggregated vertices not "definitively" assigned to processors are
// arbitrated by ArbitrateAndCommunicate(). There is some
// additional logic to prevent losing root nodes in arbitration.
{
ArrayRCP<const LO> vertex2AggId = aggregates.GetVertex2AggId()->getData(0);
ArrayRCP<const LO> procWinner = aggregates.GetProcWinner()->getData(0);
ArrayRCP<double> weights = distWeights->getDataNonConst(0);
for (size_t i=0;i<nonUniqueMap->getNodeNumElements();i++) {
if (procWinner[i] == MUELU_UNASSIGNED) {
if (vertex2AggId[i] != MUELU_UNAGGREGATED) {
weights[i] = 1.;
if (aggregates.IsRoot(i)) weights[i] = 2.;
}
}
}
// views on distributed vectors are freed here.
}
//TODO JJH We want to skip this call
myWidget.ArbitrateAndCommunicate(*distWeights, aggregates, true);
// All tentatively assigned vertices are now definitive
// Tentatively assign any vertex (ghost or local) which neighbors a root
// to the aggregate associated with the root.
{
ArrayRCP<LO> vertex2AggId = aggregates.GetVertex2AggId()->getDataNonConst(0);
ArrayRCP<const LO> procWinner = aggregates.GetProcWinner()->getData(0);
ArrayRCP<double> weights = distWeights->getDataNonConst(0);
for (my_size_t i = 0; i < nVertices; i++) {
if ( aggregates.IsRoot(i) && (procWinner[i] == myPid) ) {
// neighOfINode is the neighbor node list of node 'i'.
ArrayView<const LO> neighOfINode = graph.getNeighborVertices(i);
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int colj = *it;
if (vertex2AggId[colj] == MUELU_UNAGGREGATED) {
weights[colj]= 1.;
vertex2AggId[colj] = vertex2AggId[i];
}
}
}
}
// views on distributed vectors are freed here.
}
//TODO JJH We want to skip this call
myWidget.ArbitrateAndCommunicate(*distWeights, aggregates, true);
// All tentatively assigned vertices are now definitive
// Record the number of aggregated vertices
GO total_phase_one_aggregated = 0;
{
ArrayRCP<LO> vertex2AggId = aggregates.GetVertex2AggId()->getDataNonConst(0);
GO phase_one_aggregated = 0;
for (my_size_t i = 0; i < nVertices; i++) {
if (vertex2AggId[i] != MUELU_UNAGGREGATED)
phase_one_aggregated++;
}
sumAll(graph.GetComm(), phase_one_aggregated, total_phase_one_aggregated);
GO local_nVertices = nVertices, total_nVertices = 0;
sumAll(graph.GetComm(), local_nVertices, total_nVertices);
/* Among unaggregated points, see if we can make a reasonable size */
/* aggregate out of it. We do this by looking at neighbors and seeing */
/* how many are unaggregated and on my processor. Loosely, */
/* base the number of new aggregates created on the percentage of */
/* unaggregated nodes. */
ArrayRCP<double> weights = distWeights->getDataNonConst(0);
double factor = 1.;
factor = ((double) total_phase_one_aggregated)/((double)(total_nVertices + 1));
factor = pow(factor, GetPhase3AggCreation());
for (my_size_t i = 0; i < nVertices; i++) {
if (vertex2AggId[i] == MUELU_UNAGGREGATED)
{
// neighOfINode is the neighbor node list of node 'iNode'.
ArrayView<const LO> neighOfINode = graph.getNeighborVertices(i);
int rowi_N = neighOfINode.size();
int nonaggd_neighbors = 0;
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int colj = *it;
if (vertex2AggId[colj] == MUELU_UNAGGREGATED && colj < nVertices)
nonaggd_neighbors++;
}
if ( (nonaggd_neighbors > minNodesPerAggregate) &&
(((double) nonaggd_neighbors)/((double) rowi_N) > factor))
{
vertex2AggId[i] = (nAggregates)++;
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int colj = *it;
if (vertex2AggId[colj]==MUELU_UNAGGREGATED) {
vertex2AggId[colj] = vertex2AggId[i];
if (colj < nVertices) weights[colj] = 2.;
else weights[colj] = 1.;
}
}
aggregates.SetIsRoot(i);
weights[i] = 2.;
}
}
} // for (i = 0; i < nVertices; i++)
// views on distributed vectors are freed here.
}
//TODO JJH We want to skip this call
myWidget.ArbitrateAndCommunicate(*distWeights, aggregates, true);
//All tentatively assigned vertices are now definitive
if (IsPrint(Statistics1)) {
GO Nphase1_agg = nAggregates;
GO total_aggs;
sumAll(graph.GetComm(), Nphase1_agg, total_aggs);
GetOStream(Statistics1, 0) << "Phase 1 - nodes aggregated = " << total_phase_one_aggregated << std::endl;
GetOStream(Statistics1, 0) << "Phase 1 - total aggregates = " << total_aggs << std::endl;
GO i = nAggregates - Nphase1_agg;
{ GO ii; sumAll(graph.GetComm(),i,ii); i = ii; }
GetOStream(Statistics1, 0) << "Phase 3 - additional aggregates = " << i << std::endl;
}
// Determine vertices that are not shared by setting Temp to all ones
// and doing NonUnique2NonUnique(..., ADD). This sums values of all
// local copies associated with each Gid. Thus, sums > 1 are shared.
// std::cout << "exp_nrows=" << exp_nRows << " (nVertices= " << nVertices << ", numGhost=" << graph.GetNodeNumGhost() << ")" << std::endl;
// std::cout << "nonUniqueMap=" << nonUniqueMap->getNodeNumElements() << std::endl;
RCP<Xpetra::Vector<double,LO,GO,NO> > temp_ = Xpetra::VectorFactory<double,LO,GO,NO> ::Build(nonUniqueMap,false); //no need to zero out vector in ctor
temp_->putScalar(1.);
RCP<Xpetra::Vector<double,LO,GO,NO> > tempOutput_ = Xpetra::VectorFactory<double,LO,GO,NO> ::Build(nonUniqueMap);
myWidget.NonUnique2NonUnique(*temp_, *tempOutput_, Xpetra::ADD);
std::vector<bool> gidNotShared(exp_nRows);
{
ArrayRCP<const double> tempOutput = tempOutput_->getData(0);
for (int i = 0; i < exp_nRows; i++) {
if (tempOutput[i] > 1.)
gidNotShared[i] = false;
else
gidNotShared[i] = true;
}
}
// Phase 4.
double nAggregatesTarget;
nAggregatesTarget = ((double) uniqueMap->getGlobalNumElements())* (((double) uniqueMap->getGlobalNumElements())/ ((double) graph.GetGlobalNumEdges()));
GO nAggregatesLocal=nAggregates, nAggregatesGlobal; sumAll(graph.GetComm(), nAggregatesLocal, nAggregatesGlobal);
LO minNAggs; minAll(graph.GetComm(), nAggregates, minNAggs);
LO maxNAggs; maxAll(graph.GetComm(), nAggregates, maxNAggs);
//
// Only do this phase if things look really bad. THIS
// CODE IS PRETTY EXPERIMENTAL
//
#define MUELU_PHASE4BUCKETS 6
if ((nAggregatesGlobal < graph.GetComm()->getSize()) &&
(2.5*nAggregatesGlobal < nAggregatesTarget) &&
(minNAggs ==0) && (maxNAggs <= 1)) {
// Modify seed of the random algorithm used by temp_->randomize()
{
typedef Teuchos::ScalarTraits<double> scalarTrait; // temp_ is of type double.
scalarTrait::seedrandom(static_cast<unsigned int>(myPid*2 + (int) (11*scalarTrait::random())));
int k = (int)ceil( (10.*myPid)/graph.GetComm()->getSize());
for (int i = 0; i < k+7; i++) scalarTrait::random();
temp_->setSeed(static_cast<unsigned int>(scalarTrait::random()));
}
temp_->randomize();
ArrayRCP<double> temp = temp_->getDataNonConst(0);
// build a list of candidate root nodes (vertices not adjacent
// to aggregated vertices)
my_size_t nCandidates = 0;
global_size_t nCandidatesGlobal;
ArrayRCP<LO> candidates = Teuchos::arcp<LO>(nVertices+1);
double priorThreshold = 0.;
for (int kkk = 0; kkk < MUELU_PHASE4BUCKETS; kkk++) {
{
ArrayRCP<const LO> vertex2AggId = aggregates.GetVertex2AggId()->getData(0);
ArrayView<const LO> vertex2AggIdView = vertex2AggId();
RootCandidates(nVertices, vertex2AggIdView, graph, candidates, nCandidates, nCandidatesGlobal);
// views on distributed vectors are freed here.
}
double nTargetNewGuys = nAggregatesTarget - nAggregatesGlobal;
double threshold = priorThreshold + (1. - priorThreshold)*nTargetNewGuys/(nCandidatesGlobal + .001);
threshold = (threshold*(kkk+1.))/((double) MUELU_PHASE4BUCKETS);
priorThreshold = threshold;
{
ArrayRCP<LO> vertex2AggId = aggregates.GetVertex2AggId()->getDataNonConst(0);
ArrayRCP<double> weights = distWeights->getDataNonConst(0);
for (int k = 0; k < nCandidates; k++ ) {
int i = candidates[k];
if ((vertex2AggId[i] == MUELU_UNAGGREGATED) && (fabs(temp[i]) < threshold)) {
// Note: priorThreshold <= fabs(temp[i]) <= 1
// neighOfINode is the neighbor node list of node 'iNode'.
ArrayView<const LO> neighOfINode = graph.getNeighborVertices(i);
if (neighOfINode.size() > minNodesPerAggregate) { //TODO: check if this test is exactly was we want to do
int count = 0;
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int Adjacent = *it;
// This might not be true if someone close to i
// is chosen as a root via fabs(temp[]) < Threshold
if (vertex2AggId[Adjacent] == MUELU_UNAGGREGATED){
count++;
vertex2AggId[Adjacent] = nAggregates;
weights[Adjacent] = 1.;
}
}
if (count >= minNodesPerAggregate) {
vertex2AggId[i] = nAggregates++;
weights[i] = 2.;
aggregates.SetIsRoot(i);
}
else { // undo things
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int Adjacent = *it;
if (vertex2AggId[Adjacent] == nAggregates){
vertex2AggId[Adjacent] = MUELU_UNAGGREGATED;
weights[Adjacent] = 0.;
}
}
}
}
}
}
// views on distributed vectors are freed here.
}
//TODO JJH We want to skip this call
myWidget.ArbitrateAndCommunicate(*distWeights, aggregates, true);
// All tentatively assigned vertices are now definitive
nAggregatesLocal=nAggregates;
sumAll(graph.GetComm(), nAggregatesLocal, nAggregatesGlobal);
// check that there are no aggregates sizes below minNodesPerAggregate
aggregates.SetNumAggregates(nAggregates);
RemoveSmallAggs(aggregates, minNodesPerAggregate, distWeights, myWidget);
nAggregates = aggregates.GetNumAggregates();
} // one possibility
}
// Initialize things for Phase 5. This includes building the transpose
// of the matrix ONLY for transposed rows that correspond to unaggregted
// ghost vertices. Further, the transpose is only a local transpose.
// Nonzero edges which exist on other processors are not represented.
int observedNAgg=-1; //number of aggregates that contain vertices on this process
{
ArrayRCP<LO> vertex2AggId = aggregates.GetVertex2AggId()->getDataNonConst(0);
ArrayRCP<const LO> procWinner = aggregates.GetProcWinner()->getData(0);
for(LO k = 0; k < vertex2AggId.size(); ++k )
if(vertex2AggId[k]>observedNAgg)
observedNAgg=vertex2AggId[k];
observedNAgg++;
}
ArrayRCP<int> Mark = Teuchos::arcp<int>(exp_nRows+1);
ArrayRCP<int> agg_incremented = Teuchos::arcp<int>(observedNAgg);
ArrayRCP<int> SumOfMarks = Teuchos::arcp<int>(observedNAgg);
for (int i = 0; i < exp_nRows; i++) Mark[i] = MUELU_DISTONE_VERTEX_WEIGHT;
for (int i = 0; i < agg_incremented.size(); i++) agg_incremented[i] = 0;
for (int i = 0; i < SumOfMarks.size(); i++) SumOfMarks[i] = 0;
// Grab the transpose matrix graph for unaggregated ghost vertices.
// a) count the number of nonzeros per row in the transpose
std::vector<int> RowPtr(exp_nRows+1-nVertices);
//{
ArrayRCP<const LO> vertex2AggIdCst = aggregates.GetVertex2AggId()->getData(0);
for (int i = nVertices; i < exp_nRows; i++) RowPtr[i-nVertices] = 0;
for (int i = 0; i < nVertices; i++) {
// neighOfINode is the neighbor node list of node 'iNode'.
ArrayView<const LO> neighOfINode = graph.getNeighborVertices(i);
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int j = *it;
if ( (j >= nVertices) && (vertex2AggIdCst[j] == MUELU_UNAGGREGATED)){
RowPtr[j-nVertices]++;
}
}
}
// b) Convert RowPtr[i] to point to 1st first nnz spot in row i.
int iSum = 0, iTemp;
for (int i = nVertices; i < exp_nRows; i++) {
iTemp = RowPtr[i-nVertices];
RowPtr[i-nVertices] = iSum;
iSum += iTemp;
}
RowPtr[exp_nRows-nVertices] = iSum;
std::vector<LO> cols(iSum+1);
// c) Traverse matrix and insert entries in proper location.
for (int i = 0; i < nVertices; i++) {
// neighOfINode is the neighbor node list of node 'iNode'.
ArrayView<const LO> neighOfINode = graph.getNeighborVertices(i);
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int j = *it;
if ( (j >= nVertices) && (vertex2AggIdCst[j] == MUELU_UNAGGREGATED)){
cols[RowPtr[j-nVertices]++] = i;
}
}
}
// d) RowPtr[i] points to beginning of row i+1 so shift by one location.
for (int i = exp_nRows; i > nVertices; i--)
RowPtr[i-nVertices] = RowPtr[i-1-nVertices];
RowPtr[0] = 0;
// views on distributed vectors are freed here.
vertex2AggIdCst = Teuchos::null;
//}
int bestScoreCutoff;
int thresholds[10] = {300,200,100,50,25,13,7,4,2,0};
// Stick unaggregated vertices into existing aggregates as described above.
{
int ncalls=0;
for (int kk = 0; kk < 10; kk += 2) {
bestScoreCutoff = thresholds[kk];
ArrayRCP<LO> vertex2AggId = aggregates.GetVertex2AggId()->getDataNonConst(0);
ArrayRCP<const LO> procWinner = aggregates.GetProcWinner()->getData(0);
ArrayRCP<double> weights = distWeights->getDataNonConst(0);
for (int i = 0; i < exp_nRows; i++) {
if (vertex2AggId[i] == MUELU_UNAGGREGATED) {
// neighOfINode is the neighbor node list of node 'iNode'.
ArrayView<const LO> neighOfINode;
// Grab neighboring vertices which is either in graph for local ids
// or sits in transposed fragment just constructed above for ghosts.
if (i < nVertices) {
neighOfINode = graph.getNeighborVertices(i);
}
else {
LO *rowi_col = NULL, rowi_N;
rowi_col = &(cols[RowPtr[i-nVertices]]);
rowi_N = RowPtr[i+1-nVertices] - RowPtr[i-nVertices];
neighOfINode = ArrayView<const LO>(rowi_col, rowi_N);
}
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int Adjacent = *it;
int AdjacentAgg = vertex2AggId[Adjacent];
//Adjacent is aggregated and either I own the aggregate
// or I could own the aggregate after arbitration.
if ((AdjacentAgg != MUELU_UNAGGREGATED) &&
((procWinner[Adjacent] == myPid) ||
(procWinner[Adjacent] == MUELU_UNASSIGNED))){
SumOfMarks[AdjacentAgg] += Mark[Adjacent];
}
}
int best_score = MUELU_NOSCORE;
int best_agg = -1;
int BestMark = -1;
bool cannotLoseAllFriends=false; // Used to address possible loss of vertices in arbitration of shared nodes discussed above. (Initialized to false only to avoid a compiler warning).
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int Adjacent = *it;
int AdjacentAgg = vertex2AggId[Adjacent];
//Adjacent is unaggregated, has some value and no
//other processor has definitively claimed him
if ((AdjacentAgg != MUELU_UNAGGREGATED) &&
(SumOfMarks[AdjacentAgg] != 0) &&
((procWinner[Adjacent] == myPid) ||
(procWinner[Adjacent] == MUELU_UNASSIGNED ))) {
// first figure out the penalty associated with
// AdjacentAgg having already been incremented
// during this phase, then compute score.
double penalty = (double) (INCR_SCALING*agg_incremented[AdjacentAgg]);
if (penalty > MUELU_PENALTYFACTOR*((double)SumOfMarks[AdjacentAgg]))
penalty = MUELU_PENALTYFACTOR*((double)SumOfMarks[AdjacentAgg]);
int score = SumOfMarks[AdjacentAgg]- ((int) floor(penalty));
if (score > best_score) {
best_agg = AdjacentAgg;
best_score = score;
BestMark = Mark[Adjacent];
cannotLoseAllFriends = false;
// This address issue mentioned above by checking whether
// Adjacent could be lost in arbitration. weight==0 means that
// Adjacent was not set during this loop of Phase 5 (and so it
// has already undergone arbitration). GidNotShared == true
// obviously implies that Adjacent cannot be lost to arbitration
if ((weights[Adjacent]== 0.) || (gidNotShared[Adjacent] == true))
cannotLoseAllFriends = true;
}
// Another vertex within current best aggregate found.
// We should have (best_score == score). We need to see
// if we can improve BestMark and cannotLoseAllFriends.
else if (best_agg == AdjacentAgg) {
if ((weights[Adjacent]== 0.) || (gidNotShared[Adjacent] == true))
cannotLoseAllFriends = true;
if (Mark[Adjacent] > BestMark) BestMark = Mark[Adjacent];
}
}
}
// Clean up
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int Adjacent = *it;
int AdjacentAgg = vertex2AggId[Adjacent];
if (AdjacentAgg >= 0) SumOfMarks[AdjacentAgg] = 0;
}
// Tentatively assign vertex to best_agg.
if ( (best_score >= bestScoreCutoff) && (cannotLoseAllFriends)) {
TEUCHOS_TEST_FOR_EXCEPTION(best_agg == -1 || BestMark == -1, MueLu::Exceptions::RuntimeError, "MueLu::CoupledAggregationFactory internal error"); // should never happen
vertex2AggId[i] = best_agg;
weights[i] = best_score;
agg_incremented[best_agg]++;
Mark[i] = (int) ceil( ((double) BestMark)/2.);
}
}
// views on distributed vectors are freed here.
}
vertex2AggId = Teuchos::null;
procWinner = Teuchos::null;
weights = Teuchos::null;
++ncalls;
//TODO JJH We want to skip this call
myWidget.ArbitrateAndCommunicate(*distWeights, aggregates, true);
// All tentatively assigned vertices are now definitive
}
// if (graph.GetComm()->getRank()==0)
// std::cout << "#calls to Arb&Comm=" << ncalls << std::endl;
}
// Phase 6: Aggregate remain unaggregated vertices and try at all costs
// to avoid small aggregates.
// One case where we can find ourselves in this situation
// is if all vertices vk adjacent to v have already been
// put in other processor's aggregates and v does not have
// a direct connection to a local vertex in any of these
// aggregates.
int Nleftover = 0, Nsingle = 0;
{
ArrayRCP<LO> vertex2AggId = aggregates.GetVertex2AggId()->getDataNonConst(0);
ArrayRCP<double> weights = distWeights->getDataNonConst(0);
ArrayRCP<const LO> procWinner = aggregates.GetProcWinner()->getData(0);
int count = 0;
for (my_size_t i = 0; i < nVertices; i++) {
if (vertex2AggId[i] == MUELU_UNAGGREGATED) {
Nleftover++;
// neighOfINode is the neighbor node list of node 'iNode'.
ArrayView<const LO> neighOfINode = graph.getNeighborVertices(i);
// We don't want too small of an aggregate. So lets see if there is an
// unaggregated neighbor that we can also put with this vertex
vertex2AggId[i] = nAggregates;
weights[i] = 1.;
if (count == 0) aggregates.SetIsRoot(i);
count++;
for (typename ArrayView<const LO>::const_iterator it = neighOfINode.begin(); it != neighOfINode.end(); ++it) {
int j = *it;
if ((j != i)&&(vertex2AggId[j] == MUELU_UNAGGREGATED)&&
(j < nVertices)) {
vertex2AggId[j] = nAggregates;
weights[j] = 1.;
count++;
}
}
if ( count >= minNodesPerAggregate) {
nAggregates++;
count = 0;
}
}
}
// We have something which is under minNodesPerAggregate when
if (count != 0) {
#ifdef FIXME
// Can stick small aggregate with 0th aggregate?
if (nAggregates > 0) {
for (my_size_t i = 0; i < nVertices; i++) {
if ((vertex2AggId[i] == nAggregates) && (procWinner[i] == myPid)) {
vertex2AggId[i] = 0;
aggregates.SetIsRoot(i,false);
}
}
}
else {
Nsingle++;
nAggregates++;
}
#else
// Can stick small aggregate with 0th aggregate?
if (nAggregates > 0) {
for (my_size_t i = 0; i < nVertices; i++) {
// TW: This is not a real fix. This may produce ugly bad aggregates!
// I removed the procWinner[i] == myPid check. it makes no sense to me since
// it leaves vertex2AggId[i] == nAggregates -> crash in ComputeAggregateSizes().
// Maybe it's better to add the leftovers to the last generated agg on the current proc.
// The best solution would be to add them to the "next"/nearest aggregate, that may be
// on an other processor
if (vertex2AggId[i] == nAggregates) {
vertex2AggId[i] = nAggregates-1; //0;
aggregates.SetIsRoot(i,false);
}
}
}
else {
Nsingle++;
nAggregates++;
}
#endif
}
// views on distributed vectors are freed here.
}
//TODO JJH We want to skip this call
myWidget.ArbitrateAndCommunicate(*distWeights, aggregates, false);
if (IsPrint(Statistics1)) {
GO total_Nsingle=0; sumAll(graph.GetComm(), (GO)Nsingle, total_Nsingle);
GO total_Nleftover=0; sumAll(graph.GetComm(), (GO)Nleftover, total_Nleftover);
// GO total_aggs; sumAll(graph.GetComm(), (GO)nAggregates, total_aggs);
// GetOStream(Statistics1, 0) << "Phase 6 - total aggregates = " << total_aggs << std::endl;
GetOStream(Statistics1, 0) << "Phase 6 - leftovers = " << total_Nleftover << " and singletons = " << total_Nsingle << std::endl;
}
aggregates.SetNumAggregates(nAggregates);
} //AggregateLeftovers
void AlgebraicPermutationStrategy<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::BuildPermutation(const Teuchos::RCP<Matrix> & A, const Teuchos::RCP<const Map> permRowMap, Level & currentLevel, const FactoryBase* genFactory) const {
#ifndef HAVE_MUELU_INST_COMPLEX_INT_INT
const Teuchos::RCP< const Teuchos::Comm< int > > comm = A->getRowMap()->getComm();
int numProcs = comm->getSize();
int myRank = comm->getRank();
/*if( permRowMap == Teuchos::null ) {
permRowMap = A->getRowMap(); // use full row map of A
}*/
size_t nDofsPerNode = 1;
if (A->IsView("stridedMaps")) {
Teuchos::RCP<const Map> permRowMapStrided = A->getRowMap("stridedMaps");
nDofsPerNode = Teuchos::rcp_dynamic_cast<const StridedMap>(permRowMapStrided)->getFixedBlockSize();
}
//GetOStream(Runtime0, 0) << "Perform generation of permutation operators on " << mapName_ << " map with " << permRowMap->getGlobalNumElements() << " elements" << std::endl;
std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> > permutedDiagCandidates;
std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> > keepDiagonalEntries;
std::vector<Scalar> Weights;
// loop over all local rows in matrix A and keep diagonal entries if corresponding
// matrix rows are not contained in permRowMap
for (size_t row = 0; row < A->getRowMap()->getNodeNumElements(); row++) {
GlobalOrdinal grow = A->getRowMap()->getGlobalElement(row);
if(permRowMap->isNodeGlobalElement(grow) == true) continue;
size_t nnz = A->getNumEntriesInLocalRow(row);
// extract local row information from matrix
Teuchos::ArrayView<const LocalOrdinal> indices;
Teuchos::ArrayView<const Scalar> vals;
A->getLocalRowView(row, indices, vals);
TEUCHOS_TEST_FOR_EXCEPTION(Teuchos::as<size_t>(indices.size()) != nnz, Exceptions::RuntimeError, "MueLu::PermutationFactory::Build: number of nonzeros not equal to number of indices? Error.");
// find column entry with max absolute value
GlobalOrdinal gMaxValIdx = 0;
Scalar norm1 = 0.0;
Scalar maxVal = 0.0;
for (size_t j = 0; j < Teuchos::as<size_t>(indices.size()); j++) {
norm1 += std::abs(vals[j]);
if(std::abs(vals[j]) > maxVal) {
maxVal = std::abs(vals[j]);
gMaxValIdx = A->getColMap()->getGlobalElement(indices[j]);
}
}
if(grow == gMaxValIdx) // only keep row/col pair if it's diagonal dominant!!!
keepDiagonalEntries.push_back(std::make_pair(grow,grow));
}
//////////
// handle rows that are marked to be relevant for permutations
for (size_t row = 0; row < permRowMap->getNodeNumElements(); row++) {
GlobalOrdinal grow = permRowMap->getGlobalElement(row);
LocalOrdinal lArow = A->getRowMap()->getLocalElement(grow);
size_t nnz = A->getNumEntriesInLocalRow(lArow);
// extract local row information from matrix
Teuchos::ArrayView<const LocalOrdinal> indices;
Teuchos::ArrayView<const Scalar> vals;
A->getLocalRowView(lArow, indices, vals);
TEUCHOS_TEST_FOR_EXCEPTION(Teuchos::as<size_t>(indices.size()) != nnz, Exceptions::RuntimeError, "MueLu::PermutationFactory::Build: number of nonzeros not equal to number of indices? Error.");
// find column entry with max absolute value
GlobalOrdinal gMaxValIdx = 0;
Scalar norm1 = 0.0;
Scalar maxVal = 0.0;
for (size_t j = 0; j < Teuchos::as<size_t>(indices.size()); j++) {
norm1 += std::abs(vals[j]);
if(std::abs(vals[j]) > maxVal) {
maxVal = std::abs(vals[j]);
gMaxValIdx = A->getColMap()->getGlobalElement(indices[j]);
}
}
if(std::abs(maxVal) > 0.0) { // keep only max Entries \neq 0.0
permutedDiagCandidates.push_back(std::make_pair(grow,gMaxValIdx));
Weights.push_back(maxVal/(norm1*Teuchos::as<Scalar>(nnz)));
} else {
std::cout << "ATTENTION: row " << grow << " has only zero entries -> singular matrix!" << std::endl;
}
}
// sort Weights in descending order
std::vector<int> permutation;
sortingPermutation(Weights,permutation);
// create new vector with exactly one possible entry for each column
// each processor which requests the global column id gcid adds 1 to gColVec
// gColVec will be summed up over all processors and communicated to gDomVec
// which is based on the non-overlapping domain map of A.
Teuchos::RCP<Vector> gColVec = VectorFactory::Build(A->getColMap());
Teuchos::RCP<Vector> gDomVec = VectorFactory::Build(A->getDomainMap());
gColVec->putScalar(0.0);
gDomVec->putScalar(0.0);
// put in all keep diagonal entries
for (typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::const_iterator p = keepDiagonalEntries.begin(); p != keepDiagonalEntries.end(); ++p) {
gColVec->sumIntoGlobalValue((*p).second,1.0);
}
Teuchos::RCP<Export> exporter = ExportFactory::Build(gColVec->getMap(), gDomVec->getMap());
gDomVec->doExport(*gColVec,*exporter,Xpetra::ADD); // communicate blocked gcolids to all procs
gColVec->doImport(*gDomVec,*exporter,Xpetra::INSERT);
std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> > permutedDiagCandidatesFiltered; // TODO reserve memory
std::map<GlobalOrdinal, Scalar> gColId2Weight;
Teuchos::ArrayRCP< Scalar > ddata = gColVec->getDataNonConst(0);
for(size_t i = 0; i < permutedDiagCandidates.size(); ++i) {
// loop over all candidates
std::pair<GlobalOrdinal, GlobalOrdinal> pp = permutedDiagCandidates[permutation[i]];
GlobalOrdinal grow = pp.first;
GlobalOrdinal gcol = pp.second;
LocalOrdinal lcol = A->getColMap()->getLocalElement(gcol);
//Teuchos::ArrayRCP< Scalar > ddata = gColVec->getDataNonConst(0);
if(ddata[lcol] > 0.0){
continue; // skip lcol: column already handled by another row
}
// mark column as already taken
ddata[lcol]++;
permutedDiagCandidatesFiltered.push_back(std::make_pair(grow,gcol));
gColId2Weight[gcol] = Weights[permutation[i]];
}
// communicate how often each column index is requested by the different procs
gDomVec->doExport(*gColVec,*exporter,Xpetra::ADD);
gColVec->doImport(*gDomVec,*exporter,Xpetra::INSERT); // probably not needed // TODO check me
//*****************************************************************************************
// first communicate ALL global ids of column indices which are requested by more
// than one proc to all other procs
// detect which global col indices are requested by more than one proc
// and store them in the multipleColRequests vector
std::vector<GlobalOrdinal> multipleColRequests; // store all global column indices from current processor that are also
// requested by another processor. This is possible, since they are stored
// in gDomVec which is based on the nonoverlapping domain map. That is, each
// global col id is handled by exactly one proc.
std::queue<GlobalOrdinal> unusedColIdx; // unused column indices on current processor
for(size_t sz = 0; sz<gDomVec->getLocalLength(); ++sz) {
Teuchos::ArrayRCP< const Scalar > arrDomVec = gDomVec->getData(0);
if(arrDomVec[sz] > 1.0) {
multipleColRequests.push_back(gDomVec->getMap()->getGlobalElement(sz));
} else if(arrDomVec[sz] == 0.0) {
unusedColIdx.push(gDomVec->getMap()->getGlobalElement(sz));
}
}
// communicate the global number of column indices which are requested by more than one proc
LocalOrdinal localMultColRequests = Teuchos::as<LocalOrdinal>(multipleColRequests.size());
LocalOrdinal globalMultColRequests = 0;
// sum up all entries in multipleColRequests over all processors
sumAll(gDomVec->getMap()->getComm(), (LocalOrdinal)localMultColRequests, globalMultColRequests);
if(globalMultColRequests > 0) {
// special handling: two processors request the same global column id.
// decide which processor gets it
// distribute number of multipleColRequests to all processors
// each processor stores how many column ids for exchange are handled by the cur proc
std::vector<GlobalOrdinal> numMyMultColRequests(numProcs,0);
std::vector<GlobalOrdinal> numGlobalMultColRequests(numProcs,0);
numMyMultColRequests[myRank] = localMultColRequests;
Teuchos::reduceAll(*comm,Teuchos::REDUCE_MAX,numProcs,&numMyMultColRequests[0],&numGlobalMultColRequests[0]);
// communicate multipleColRequests entries to all processors
int nMyOffset = 0;
for (int i=0; i<myRank-1; i++)
nMyOffset += numGlobalMultColRequests[i]; // calculate offset to store the weights on the corresponding place in procOverlappingWeights
GlobalOrdinal zero=0;
std::vector<GlobalOrdinal> procMultRequestedColIds(globalMultColRequests,zero);
std::vector<GlobalOrdinal> global_procMultRequestedColIds(globalMultColRequests,zero);
// loop over all local column GIDs that are also requested by other procs
for(size_t i = 0; i < multipleColRequests.size(); i++) {
procMultRequestedColIds[nMyOffset + i] = multipleColRequests[i]; // all weights are > 0 ?
}
// template ordinal, package (double)
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, Teuchos::as<int>(globalMultColRequests), &procMultRequestedColIds[0], &global_procMultRequestedColIds[0]);
// loop over global_procOverlappingWeights and eliminate wrong entries...
for (size_t k = 0; k<global_procMultRequestedColIds.size(); k++) {
GlobalOrdinal globColId = global_procMultRequestedColIds[k];
std::vector<Scalar> MyWeightForColId(numProcs,0);
std::vector<Scalar> GlobalWeightForColId(numProcs,0);
if(gColVec->getMap()->isNodeGlobalElement(globColId)) {
MyWeightForColId[myRank] = gColId2Weight[globColId];
} else {
MyWeightForColId[myRank] = 0.0;
}
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, numProcs, &MyWeightForColId[0], &GlobalWeightForColId[0]);
if(gColVec->getMap()->isNodeGlobalElement(globColId)) {
// note: 2 procs could have the same weight for a column index.
// pick the first one.
Scalar winnerValue = 0.0;
int winnerProcRank = 0;
for (int proc = 0; proc < numProcs; proc++) {
if(GlobalWeightForColId[proc] > winnerValue) {
winnerValue = GlobalWeightForColId[proc];
winnerProcRank = proc;
}
}
// winnerProcRank is the winner for handling globColId.
// winnerProcRank is unique (even if two procs have the same weight for a column index)
if(myRank != winnerProcRank) {
// remove corresponding entry from permutedDiagCandidatesFiltered
typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::iterator p = permutedDiagCandidatesFiltered.begin();
while(p != permutedDiagCandidatesFiltered.end() )
{
if((*p).second == globColId)
p = permutedDiagCandidatesFiltered.erase(p);
else
p++;
}
}
} // end if isNodeGlobalElement
} // end loop over global_procOverlappingWeights and eliminate wrong entries...
} // end if globalMultColRequests > 0
// put together all pairs:
//size_t sizeRowColPairs = keepDiagonalEntries.size() + permutedDiagCandidatesFiltered.size();
std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> > RowColPairs;
RowColPairs.insert( RowColPairs.end(), keepDiagonalEntries.begin(), keepDiagonalEntries.end());
RowColPairs.insert( RowColPairs.end(), permutedDiagCandidatesFiltered.begin(), permutedDiagCandidatesFiltered.end());
#ifdef DEBUG_OUTPUT
//&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&
// plausibility check
gColVec->putScalar(0.0);
gDomVec->putScalar(0.0);
typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::iterator pl = RowColPairs.begin();
while(pl != RowColPairs.end() )
{
//GlobalOrdinal ik = (*pl).first;
GlobalOrdinal jk = (*pl).second;
gColVec->sumIntoGlobalValue(jk,1.0);
pl++;
}
gDomVec->doExport(*gColVec,*exporter,Xpetra::ADD);
for(size_t sz = 0; sz<gDomVec->getLocalLength(); ++sz) {
Teuchos::ArrayRCP< const Scalar > arrDomVec = gDomVec->getData(0);
if(arrDomVec[sz] > 1.0) {
GetOStream(Runtime0,0) << "RowColPairs has multiple column [" << sz << "]=" << arrDomVec[sz] << std::endl;
} else if(arrDomVec[sz] == 0.0) {
GetOStream(Runtime0,0) << "RowColPairs has empty column [" << sz << "]=" << arrDomVec[sz] << std::endl;
}
}
//&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&
#endif
//////////////////////////////////////////////////
// assumption: on each processor RowColPairs now contains
// a valid set of (row,column) pairs, where the row entries
// are a subset of the processor's rows and the column entries
// are unique throughout all processors.
// Note: the RowColPairs are only defined for a subset of all rows,
// so there might be rows without an entry in RowColPairs.
// It can be, that some rows seem to be missing in RowColPairs, since
// the entry in that row with maximum absolute value has been reserved
// by another row already (e.g. as already diagonal dominant row outside
// of perRowMap).
// In fact, the RowColPairs vector only defines the (row,column) pairs
// that will be definitely moved to the diagonal after permutation.
#ifdef DEBUG_OUTPUT
// for (typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::const_iterator p = RowColPairs.begin(); p != RowColPairs.end(); ++p) {
// std::cout << "proc: " << myRank << " r/c: " << (*p).first << "/" << (*p).second << std::endl;
// }
// for (typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::const_iterator p = RowColPairs.begin(); p != RowColPairs.end(); ++p)
// {
//// if((*p).first != (*p).second) std::cout << "difference: " << (*p).first << " " << (*p).second << std::endl;
// std::cout << (*p).first +1 << " " << (*p).second+1 << std::endl;
// }
// std::cout << "\n";
#endif
// vectors to store permutation information
Teuchos::RCP<Vector> Pperm = VectorFactory::Build(A->getRowMap());
Teuchos::RCP<Vector> Qperm = VectorFactory::Build(A->getDomainMap()); // global variant (based on domain map)
Teuchos::RCP<Vector> lQperm = VectorFactory::Build(A->getColMap()); // local variant (based on column map)
Teuchos::ArrayRCP< Scalar > PpermData = Pperm->getDataNonConst(0);
Teuchos::ArrayRCP< Scalar > QpermData = Qperm->getDataNonConst(0);
Pperm->putScalar(0.0);
Qperm->putScalar(0.0);
lQperm->putScalar(0.0);
// setup exporter for Qperm
Teuchos::RCP<Export> QpermExporter = ExportFactory::Build(lQperm->getMap(), Qperm->getMap());
Teuchos::RCP<Vector> RowIdStatus = VectorFactory::Build(A->getRowMap());
Teuchos::RCP<Vector> ColIdStatus = VectorFactory::Build(A->getDomainMap()); // global variant (based on domain map)
Teuchos::RCP<Vector> lColIdStatus = VectorFactory::Build(A->getColMap()); // local variant (based on column map)
Teuchos::RCP<Vector> ColIdUsed = VectorFactory::Build(A->getDomainMap()); // mark column ids to be already in use
Teuchos::ArrayRCP< Scalar > RowIdStatusArray = RowIdStatus->getDataNonConst(0);
Teuchos::ArrayRCP< Scalar > ColIdStatusArray = ColIdStatus->getDataNonConst(0);
Teuchos::ArrayRCP< Scalar > lColIdStatusArray = lColIdStatus->getDataNonConst(0);
Teuchos::ArrayRCP< Scalar > ColIdUsedArray = ColIdUsed->getDataNonConst(0); // not sure about this
RowIdStatus->putScalar(0.0);
ColIdStatus->putScalar(0.0);
lColIdStatus->putScalar(0.0);
ColIdUsed->putScalar(0.0); // no column ids are used
// count wide-range permutations
// a wide-range permutation is defined as a permutation of rows/columns which do not
// belong to the same node
LocalOrdinal lWideRangeRowPermutations = 0;
GlobalOrdinal gWideRangeRowPermutations = 0;
LocalOrdinal lWideRangeColPermutations = 0;
GlobalOrdinal gWideRangeColPermutations = 0;
// run 1: mark all "identity" permutations
typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::iterator p = RowColPairs.begin();
while(p != RowColPairs.end() )
{
GlobalOrdinal ik = (*p).first;
GlobalOrdinal jk = (*p).second;
LocalOrdinal lik = A->getRowMap()->getLocalElement(ik);
LocalOrdinal ljk = A->getColMap()->getLocalElement(jk);
if(RowIdStatusArray[lik] == 0.0) {
RowIdStatusArray[lik] = 1.0; // use this row id
lColIdStatusArray[ljk] = 1.0; // use this column id
Pperm->replaceLocalValue(lik, ik);
lQperm->replaceLocalValue(ljk, ik); // use column map
ColIdUsed->replaceGlobalValue(ik,1.0); // ik is now used
p = RowColPairs.erase(p);
// detect wide range permutations
if(floor(ik/nDofsPerNode) != floor(jk/nDofsPerNode)) {
lWideRangeColPermutations++;
}
}
else
p++;
}
// communicate column map -> domain map
Qperm->doExport(*lQperm,*QpermExporter,Xpetra::ABSMAX);
ColIdStatus->doExport(*lColIdStatus,*QpermExporter,Xpetra::ABSMAX);
// plausibility check
if(RowColPairs.size()>0) GetOStream(Warnings0,0) << "MueLu::PermutationFactory: There are Row/Col pairs left!!!" << std::endl; // TODO fix me
// close Pperm
// count, how many row permutations are missing on current proc
size_t cntFreeRowIdx = 0;
std::queue<GlobalOrdinal> qFreeGRowIdx; // store global row ids of "free" rows
for (size_t lik = 0; lik < RowIdStatus->getLocalLength(); ++lik) {
if(RowIdStatusArray[lik] == 0.0) {
cntFreeRowIdx++;
qFreeGRowIdx.push(RowIdStatus->getMap()->getGlobalElement(lik));
}
}
// fix Pperm
for (size_t lik = 0; lik < RowIdStatus->getLocalLength(); ++lik) {
if(RowIdStatusArray[lik] == 0.0) {
RowIdStatusArray[lik] = 1.0; // use this row id
Pperm->replaceLocalValue(lik, qFreeGRowIdx.front());
// detect wide range permutations
if(floor(qFreeGRowIdx.front()/nDofsPerNode) != floor(RowIdStatus->getMap()->getGlobalElement(lik)/nDofsPerNode)) {
lWideRangeRowPermutations++;
}
qFreeGRowIdx.pop();
}
}
// close Qperm (free permutation entries in Qperm)
size_t cntFreeColIdx = 0;
std::queue<GlobalOrdinal> qFreeGColIdx; // store global column ids of "free" available columns
for (size_t ljk = 0; ljk < ColIdStatus->getLocalLength(); ++ljk) {
if(ColIdStatusArray[ljk] == 0.0) {
cntFreeColIdx++;
qFreeGColIdx.push(ColIdStatus->getMap()->getGlobalElement(ljk));
}
}
size_t cntUnusedColIdx = 0;
std::queue<GlobalOrdinal> qUnusedGColIdx; // store global column ids of "free" available columns
for (size_t ljk = 0; ljk < ColIdUsed->getLocalLength(); ++ljk) {
if(ColIdUsedArray[ljk] == 0.0) {
cntUnusedColIdx++;
qUnusedGColIdx.push(ColIdUsed->getMap()->getGlobalElement(ljk));
}
}
// fix Qperm with local entries
for (size_t ljk = 0; ljk < ColIdStatus->getLocalLength(); ++ljk) {
// stop if no (local) unused column idx are left
if(cntUnusedColIdx == 0) break;
if(ColIdStatusArray[ljk] == 0.0) {
ColIdStatusArray[ljk] = 1.0; // use this row id
Qperm->replaceLocalValue(ljk, qUnusedGColIdx.front()); // loop over ColIdStatus (lives on domain map)
ColIdUsed->replaceGlobalValue(qUnusedGColIdx.front(),1.0); // ljk is now used, too
// detect wide range permutations
if(floor(qUnusedGColIdx.front()/nDofsPerNode) != floor(ColIdStatus->getMap()->getGlobalElement(ljk)/nDofsPerNode)) {
lWideRangeColPermutations++;
}
qUnusedGColIdx.pop();
cntUnusedColIdx--;
cntFreeColIdx--;
}
}
//Qperm->doExport(*lQperm,*QpermExporter,Xpetra::ABSMAX); // no export necessary, since changes only locally
//ColIdStatus->doExport(*lColIdStatus,*QpermExporter,Xpetra::ABSMAX);
// count, how many unused column idx are needed on current processor
// to complete Qperm
cntFreeColIdx = 0;
for (size_t ljk = 0; ljk < ColIdStatus->getLocalLength(); ++ljk) { // TODO avoid this loop
if(ColIdStatusArray[ljk] == 0.0) {
cntFreeColIdx++;
}
}
GlobalOrdinal global_cntFreeColIdx = 0;
LocalOrdinal local_cntFreeColIdx = cntFreeColIdx;
sumAll(comm, Teuchos::as<GlobalOrdinal>(local_cntFreeColIdx), global_cntFreeColIdx);
#ifdef DEBUG_OUTPUT
std::cout << "global # of empty column idx entries in Qperm: " << global_cntFreeColIdx << std::endl;
#endif
// avoid global communication if possible
if(global_cntFreeColIdx > 0) {
// 1) count how many unused column ids are left
GlobalOrdinal global_cntUnusedColIdx = 0;
LocalOrdinal local_cntUnusedColIdx = cntUnusedColIdx;
sumAll(comm, Teuchos::as<GlobalOrdinal>(local_cntUnusedColIdx), global_cntUnusedColIdx);
#ifdef DEBUG_OUTPUT
std::cout << "global # of unused column idx: " << global_cntUnusedColIdx << std::endl;
#endif
// 2) communicate how many unused column ids are available on procs
std::vector<LocalOrdinal> local_UnusedColIdxOnProc (numProcs);
std::vector<LocalOrdinal> global_UnusedColIdxOnProc(numProcs);
local_UnusedColIdxOnProc[myRank] = local_cntUnusedColIdx;
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, numProcs, &local_UnusedColIdxOnProc[0], &global_UnusedColIdxOnProc[0]);
#ifdef DEBUG_OUTPUT
std::cout << "PROC " << myRank << " global num unused indices per proc: ";
for (size_t ljk = 0; ljk < global_UnusedColIdxOnProc.size(); ++ljk) {
std::cout << " " << global_UnusedColIdxOnProc[ljk];
}
std::cout << std::endl;
#endif
// 3) build array of length global_cntUnusedColIdx to globally replicate unused column idx
std::vector<GlobalOrdinal> local_UnusedColIdxVector(Teuchos::as<size_t>(global_cntUnusedColIdx));
std::vector<GlobalOrdinal> global_UnusedColIdxVector(Teuchos::as<size_t>(global_cntUnusedColIdx));
GlobalOrdinal global_cntUnusedColIdxStartIter = 0;
for(int proc=0; proc<myRank; proc++) {
global_cntUnusedColIdxStartIter += global_UnusedColIdxOnProc[proc];
}
for(GlobalOrdinal k = global_cntUnusedColIdxStartIter; k < global_cntUnusedColIdxStartIter+local_cntUnusedColIdx; k++) {
local_UnusedColIdxVector[k] = qUnusedGColIdx.front();
qUnusedGColIdx.pop();
}
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, Teuchos::as<int>(global_cntUnusedColIdx), &local_UnusedColIdxVector[0], &global_UnusedColIdxVector[0]);
#ifdef DEBUG_OUTPUT
std::cout << "PROC " << myRank << " global UnusedGColIdx: ";
for (size_t ljk = 0; ljk < global_UnusedColIdxVector.size(); ++ljk) {
std::cout << " " << global_UnusedColIdxVector[ljk];
}
std::cout << std::endl;
#endif
// 4) communicate, how many column idx are needed on each processor
// to complete Qperm
std::vector<LocalOrdinal> local_EmptyColIdxOnProc (numProcs);
std::vector<LocalOrdinal> global_EmptyColIdxOnProc(numProcs);
local_EmptyColIdxOnProc[myRank] = local_cntFreeColIdx;
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, numProcs, &local_EmptyColIdxOnProc[0], &global_EmptyColIdxOnProc[0]);
#ifdef DEBUG_OUTPUT
std::cout << "PROC " << myRank << " global num of needed column indices: ";
for (size_t ljk = 0; ljk < global_EmptyColIdxOnProc.size(); ++ljk) {
std::cout << " " << global_EmptyColIdxOnProc[ljk];
}
std::cout << std::endl;
#endif
// 5) determine first index in global_UnusedColIdxVector for unused column indices,
// that are marked to be used by this processor
GlobalOrdinal global_UnusedColStartIdx = 0;
for(int proc=0; proc<myRank; proc++) {
global_UnusedColStartIdx += global_EmptyColIdxOnProc[proc];
}
#ifdef DEBUG_OUTPUT
GetOStream(Statistics0,0) << "PROC " << myRank << " is allowd to use the following column gids: ";
for(GlobalOrdinal k = global_UnusedColStartIdx; k < global_UnusedColStartIdx + Teuchos::as<GlobalOrdinal>(cntFreeColIdx); k++) {
GetOStream(Statistics0,0) << global_UnusedColIdxVector[k] << " ";
}
GetOStream(Statistics0,0) << std::endl;
#endif
// 6.) fix Qperm with global entries
GlobalOrdinal array_iter = 0;
for (size_t ljk = 0; ljk < ColIdStatus->getLocalLength(); ++ljk) {
if(ColIdStatusArray[ljk] == 0.0) {
ColIdStatusArray[ljk] = 1.0; // use this row id
Qperm->replaceLocalValue(ljk, global_UnusedColIdxVector[global_UnusedColStartIdx + array_iter]);
ColIdUsed->replaceGlobalValue(global_UnusedColIdxVector[global_UnusedColStartIdx + array_iter],1.0);
// detect wide range permutations
if(floor(global_UnusedColIdxVector[global_UnusedColStartIdx + array_iter]/nDofsPerNode) != floor(ColIdStatus->getMap()->getGlobalElement(ljk)/nDofsPerNode)) {
lWideRangeColPermutations++;
}
array_iter++;
//cntUnusedColIdx--; // check me
}
}
} // end if global_cntFreeColIdx > 0
/////////////////// Qperm should be fine now...
// create new empty Matrix
Teuchos::RCP<CrsMatrixWrap> permPTmatrix = Teuchos::rcp(new CrsMatrixWrap(A->getRowMap(),1,Xpetra::StaticProfile));
Teuchos::RCP<CrsMatrixWrap> permQTmatrix = Teuchos::rcp(new CrsMatrixWrap(A->getRowMap(),1,Xpetra::StaticProfile));
for(size_t row=0; row<A->getNodeNumRows(); row++) {
Teuchos::ArrayRCP<GlobalOrdinal> indoutP(1,Teuchos::as<GO>(PpermData[row])); // column idx for Perm^T
Teuchos::ArrayRCP<GlobalOrdinal> indoutQ(1,Teuchos::as<GO>(QpermData[row])); // column idx for Qperm
Teuchos::ArrayRCP<Scalar> valout(1,1.0);
permPTmatrix->insertGlobalValues(A->getRowMap()->getGlobalElement(row), indoutP.view(0,indoutP.size()), valout.view(0,valout.size()));
permQTmatrix->insertGlobalValues (A->getRowMap()->getGlobalElement(row), indoutQ.view(0,indoutQ.size()), valout.view(0,valout.size()));
}
permPTmatrix->fillComplete();
permQTmatrix->fillComplete();
Teuchos::RCP<Matrix> permPmatrix = Utils2::Transpose(permPTmatrix,true);
for(size_t row=0; row<permPTmatrix->getNodeNumRows(); row++) {
if(permPTmatrix->getNumEntriesInLocalRow(row) != 1)
GetOStream(Warnings0,0) <<"#entries in row " << row << " of permPTmatrix is " << permPTmatrix->getNumEntriesInLocalRow(row) << std::endl;
if(permPmatrix->getNumEntriesInLocalRow(row) != 1)
GetOStream(Warnings0,0) <<"#entries in row " << row << " of permPmatrix is " << permPmatrix->getNumEntriesInLocalRow(row) << std::endl;
if(permQTmatrix->getNumEntriesInLocalRow(row) != 1)
GetOStream(Warnings0,0) <<"#entries in row " << row << " of permQmatrix is " << permQTmatrix->getNumEntriesInLocalRow(row) << std::endl;
}
// build permP * A * permQT
Teuchos::RCP<Matrix> ApermQt = Utils::Multiply(*A, false, *permQTmatrix, false);
Teuchos::RCP<Matrix> permPApermQt = Utils::Multiply(*permPmatrix, false, *ApermQt, false);
/*
MueLu::Utils<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Write("A.mat", *A);
MueLu::Utils<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Write("permP.mat", *permPmatrix);
MueLu::Utils<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Write("permQt.mat", *permQTmatrix);
MueLu::Utils<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Write("permPApermQt.mat", *permPApermQt);
*/
// build scaling matrix
Teuchos::RCP<Vector> diagVec = VectorFactory::Build(permPApermQt->getRowMap(),true);
Teuchos::RCP<Vector> invDiagVec = VectorFactory::Build(permPApermQt->getRowMap(),true);
Teuchos::ArrayRCP< const Scalar > diagVecData = diagVec->getData(0);
Teuchos::ArrayRCP< Scalar > invDiagVecData = invDiagVec->getDataNonConst(0);
permPApermQt->getLocalDiagCopy(*diagVec);
for(size_t i = 0; i<diagVec->getMap()->getNodeNumElements(); ++i) {
if(diagVecData[i] != 0.0)
invDiagVecData[i] = 1/diagVecData[i];
else {
invDiagVecData[i] = 1.0;
GetOStream(Statistics0,0) << "MueLu::PermutationFactory: found zero on diagonal in row " << i << std::endl;
}
}
Teuchos::RCP<CrsMatrixWrap> diagScalingOp = Teuchos::rcp(new CrsMatrixWrap(permPApermQt->getRowMap(),1,Xpetra::StaticProfile));
for(size_t row=0; row<A->getNodeNumRows(); row++) {
Teuchos::ArrayRCP<GlobalOrdinal> indout(1,permPApermQt->getRowMap()->getGlobalElement(row)); // column idx for Perm^T
Teuchos::ArrayRCP<Scalar> valout(1,invDiagVecData[row]);
diagScalingOp->insertGlobalValues(A->getRowMap()->getGlobalElement(row), indout.view(0,indout.size()), valout.view(0,valout.size()));
}
diagScalingOp->fillComplete();
Teuchos::RCP<Matrix> scaledA = Utils::Multiply(*diagScalingOp, false, *permPApermQt, false);
currentLevel.Set("A", Teuchos::rcp_dynamic_cast<Matrix>(scaledA), genFactory/*this*/);
currentLevel.Set("permA", Teuchos::rcp_dynamic_cast<Matrix>(permPApermQt), genFactory/*this*/); // TODO careful with this!!!
currentLevel.Set("permP", Teuchos::rcp_dynamic_cast<Matrix>(permPmatrix), genFactory/*this*/);
currentLevel.Set("permQT", Teuchos::rcp_dynamic_cast<Matrix>(permQTmatrix), genFactory/*this*/);
currentLevel.Set("permScaling", Teuchos::rcp_dynamic_cast<Matrix>(diagScalingOp), genFactory/*this*/);
//// count row permutations
// count zeros on diagonal in P -> number of row permutations
Teuchos::RCP<Vector> diagPVec = VectorFactory::Build(permPmatrix->getRowMap(),true);
permPmatrix->getLocalDiagCopy(*diagPVec);
Teuchos::ArrayRCP< const Scalar > diagPVecData = diagPVec->getData(0);
LocalOrdinal lNumRowPermutations = 0;
GlobalOrdinal gNumRowPermutations = 0;
for(size_t i = 0; i<diagPVec->getMap()->getNodeNumElements(); ++i) {
if(diagPVecData[i] == 0.0) {
lNumRowPermutations++;
}
}
// sum up all entries in multipleColRequests over all processors
sumAll(diagPVec->getMap()->getComm(), Teuchos::as<GlobalOrdinal>(lNumRowPermutations), gNumRowPermutations);
//// count column permutations
// count zeros on diagonal in Q^T -> number of column permutations
Teuchos::RCP<Vector> diagQTVec = VectorFactory::Build(permQTmatrix->getRowMap(),true);
permQTmatrix->getLocalDiagCopy(*diagQTVec);
Teuchos::ArrayRCP< const Scalar > diagQTVecData = diagQTVec->getData(0);
LocalOrdinal lNumColPermutations = 0;
GlobalOrdinal gNumColPermutations = 0;
for(size_t i = 0; i<diagQTVec->getMap()->getNodeNumElements(); ++i) {
if(diagQTVecData[i] == 0.0) {
lNumColPermutations++;
}
}
// sum up all entries in multipleColRequests over all processors
sumAll(diagQTVec->getMap()->getComm(), Teuchos::as<GlobalOrdinal>(lNumColPermutations), gNumColPermutations);
currentLevel.Set("#RowPermutations", gNumRowPermutations, genFactory/*this*/);
currentLevel.Set("#ColPermutations", gNumColPermutations, genFactory/*this*/);
currentLevel.Set("#WideRangeRowPermutations", gWideRangeRowPermutations, genFactory/*this*/);
currentLevel.Set("#WideRangeColPermutations", gWideRangeColPermutations, genFactory/*this*/);
GetOStream(Statistics0, 0) << "#Row permutations/max possible permutations: " << gNumRowPermutations << "/" << diagPVec->getMap()->getGlobalNumElements() << std::endl;
GetOStream(Statistics0, 0) << "#Column permutations/max possible permutations: " << gNumColPermutations << "/" << diagQTVec->getMap()->getGlobalNumElements() << std::endl;
GetOStream(Runtime1, 0) << "#wide range row permutations: " << gWideRangeRowPermutations << " #wide range column permutations: " << gWideRangeColPermutations << std::endl;
#else
#warning PermutationFactory not compiling/working for Scalar==complex.
#endif // #ifndef HAVE_MUELU_INST_COMPLEX_INT_INT
}
Matrix& LASSO::train(Matrix& X, Matrix& Y) {
/*XNX = [X, -X];
H_G = XNX' * XNX;
D = repmat(diag(H_G), [1, n_y]);
XNXTY = XNX' * Y;
A = (X' * X + lambda * eye(p)) \ (X' * Y);*/
Matrix& XNX = horzcat(2, &X, &uminus(X));
Matrix& H_G = XNX.transpose().mtimes(XNX);
double* Q = new double[size(H_G, 1)];
for (int i = 0; i < size(H_G, 1); i++) {
Q[i] = H_G.getEntry(i, i);
}
Matrix& XNXTY = XNX.transpose().mtimes(Y);
Matrix& A = mldivide(
plus(X.transpose().mtimes(X), times(lambda, eye(p))),
X.transpose().mtimes(Y)
);
/*AA = [subplus(A); subplus(-A)];
C = -XNXTY + lambda;
Grad = C + H_G * AA;
tol = epsilon * norm(Grad);
PGrad = zeros(size(Grad));*/
Matrix& AA = vertcat(2, &subplus(A), &subplus(uminus(A)));
Matrix& C = plus(uminus(XNXTY), lambda);
Matrix& Grad = plus(C, mtimes(H_G, AA));
double tol = epsilon * norm(Grad);
Matrix& PGrad = zeros(size(Grad));
std::list<double> J;
double fval = 0;
// J(1) = sum(sum((Y - X * A).^2)) / 2 + lambda * sum(sum(abs(A)));
if (calc_OV) {
fval = sumAll(pow(minus(Y, mtimes(X, A)), 2)) / 2 +
lambda * sum(sum(abs(A)));
J.push_back(fval);
}
/*Matrix I_k = null;
Matrix I_k_com = null;*/
Matrix& I_k = Grad.copy();
double d = 0;
int k = 0;
DenseVector& SFPlusCi = *new DenseVector(AA.getColumnDimension());
Matrix& S = H_G;
Vector** SRows = null;
if (typeid(H_G) == typeid(DenseMatrix))
SRows = denseMatrix2DenseRowVectors(S);
else
SRows = sparseMatrix2SparseRowVectors(S);
Vector** CRows = null;
if (typeid(C) == typeid(DenseMatrix))
CRows = denseMatrix2DenseRowVectors(C);
else
CRows = sparseMatrix2SparseRowVectors(C);
double** FData = ((DenseMatrix&) AA).getData();
double* FRow = null;
double* pr = null;
int K = 2 * p;
while (true) {
/*I_k = Grad < 0 | AA > 0;
I_k_com = not(I_k);
PGrad(I_k) = Grad(I_k);
PGrad(I_k_com) = 0;*/
_or(I_k, lt(Grad, 0), gt(AA, 0));
Matrix& I_k_com = _not(I_k);
assign(PGrad, Grad);
logicalIndexingAssignment(PGrad, I_k_com, 0);
d = norm(PGrad, inf);
if (d < tol) {
if (verbose)
println("Converge successfully!");
break;
}
/*for i = 1:2*p
AA(i, :) = max(AA(i, :) - (C(i, :) + H_G(i, :) * AA) ./ (D(i, :)), 0);
end
A = AA(1:p,:) - AA(p+1:end,:);*/
for (int i = 0; i < K; i++) {
// SFPlusCi = SRows[i].operate(AA);
operate(SFPlusCi, *SRows[i], AA);
plusAssign(SFPlusCi, *CRows[i]);
timesAssign(SFPlusCi, 1 / Q[i]);
pr = SFPlusCi.getPr();
// F(i, :) = max(F(i, :) - (S(i, :) * F + C(i, :)) / D[i]), 0);
// F(i, :) = max(F(i, :) - SFPlusCi, 0)
FRow = FData[i];
for (int j = 0; j < AA.getColumnDimension(); j++) {
FRow[j] = max(FRow[j] - pr[j], 0);
}
}
// Grad = plus(C, mtimes(H_G, AA));
plus(Grad, C, mtimes(H_G, AA));
k = k + 1;
if (k > maxIter) {
if (verbose)
println("Maximal iterations");
break;
}
if (calc_OV) {
fval = sum(sum(pow(minus(Y, mtimes(XNX, AA)), 2))) / 2 +
lambda * sum(sum(abs(AA)));
J.push_back(fval);
}
if (k % 10 == 0 && verbose) {
if (calc_OV)
fprintf("Iter %d - ||PGrad||: %f, ofv: %f\n", k, d, J.back());
else
fprintf("Iter %d - ||PGrad||: %f\n", k, d);
}
}
Matrix& res = minus(
AA.getSubMatrix(0, p - 1, 0, ny - 1),
AA.getSubMatrix(p, 2 * p - 1, 0, ny - 1)
);
return res;
}
