static bool analyzePartitionMetrics( const ParameterList &metricsPlist,
const RCP<const Zoltan2::EvaluatePartition <basic_id_t> > &metricObject,
const RCP<const Comm<int>> &comm,
std::ostringstream &msg_stream) {
ArrayRCP<const metric_t> metrics = metricObject->getMetrics();
bool all_tests_pass = true;
zscalar_t metric_value = 0.0;
for (int i = 0; i < metrics.size(); i++) {
// print their names...
if (metricsPlist.isSublist(metrics[i].getName())) {
auto metric_plist = metricsPlist.sublist(metrics[i].getName());
// loop on tests
auto p= metric_plist.begin(); // iterator
while (p != metric_plist.end()) {
auto test_name = metric_plist.name(p);
if( metrics[i].hasMetricValue(test_name)) {
if(!MetricAnalyzer::MetricBoundsTest( metrics[i].getMetricValue(test_name),
test_name,
metric_plist.sublist(test_name),
comm,
msg_stream)) {
all_tests_pass = false;
}
} else msg_stream << "UNKNOWN TEST: " + test_name << std::endl;
++p;
}
} else {
msg_stream << "UNKNOWN METRIC: " + metrics[i].getName() << std::endl;
}
}
return all_tests_pass;
}
NonlinearSolver<double> NonlinearSolverBuilder::createSolver(const ParameterList& params)
{
if (params.isSublist("NOX Solver"))
{
return new NOXSolver(params);
}
else if (params.isSublist("Nonlinear Solver"))
{
ParameterList sub = params.sublist("Nonlinear Solver");
Array<string> names = tuple<string>("Newton Armijo Solver", "Newton-Armijo Solver", "NewtonArmijoSolver");
for (int i=0; i<names.size(); i++)
{
if (sub.isSublist(names[i]))
{
ParameterList subsub = sub.sublist(names[i]);
LinearSolver<double> linSolver;
if (subsub.isParameter("Linear Solver"))
{
string solverFile = subsub.get<string>("Linear Solver");
linSolver = LinearSolverBuilder::createSolver(solverFile);
}
else if (subsub.isSublist("Linear Solver"))
{
linSolver = LinearSolverBuilder::createSolver(subsub);
}
else
{
TEUCHOS_TEST_FOR_EXCEPTION(true, std::runtime_error,
"Nonlinear solver parameter list " << sub
<< " does not appear to specify a solver for the linear subproblems");
}
return new NewtonArmijoSolver<double>(subsub, linSolver);
}
}
}
else
{
TEUCHOS_TEST_FOR_EXCEPTION(true, std::runtime_error,
"Nonlinear solver parameter list " << params
<< " can't be parsed to find a nonlinear solver");
}
return NonlinearSolver<double>();
}
BelosSolver::BelosSolver(const ParameterList& params)
: LinearSolverBase<double>(params), pf_(), hasSolver_(false)
{
if (params.isSublist("Preconditioner"))
{
ParameterList precParams = params.sublist("Preconditioner");
pf_ = new ParameterListPreconditionerFactory(precParams);
}
}
std::queue<ParameterList> ComparisonHelper::getMetricsToCompare(const ParameterList &pList)
{
// extract all of the metrics to be tested
std::queue<ParameterList> metrics;
for(auto it = pList.begin(); it != pList.end(); ++it) {
if (pList.isSublist(it->first)) {
metrics.push(pList.sublist(it->first));
}
}
return metrics;
}
void ParameterListInterpreter<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::SetEasyParameterList(const Teuchos::ParameterList& constParamList) {
// Create a non const copy of the parameter list
// Working with a modifiable list is much much easier than with original one
ParameterList paramList = constParamList;
// Translate cycle type parameter
if (paramList.isParameter("cycle type")) {
std::map<std::string,CycleType> cycleMap;
cycleMap["V"] = VCYCLE;
cycleMap["W"] = WCYCLE;
std::string cycleType = paramList.get<std::string>("cycle type");
TEUCHOS_TEST_FOR_EXCEPTION(cycleMap.count(cycleType) == 0, Exceptions::RuntimeError, "Invalid cycle type: \"" << cycleType << "\"");
Cycle_ = cycleMap[cycleType];
}
this->maxCoarseSize_ = paramList.get<int> ("coarse: max size", Hierarchy::GetDefaultMaxCoarseSize());
this->numDesiredLevel_ = paramList.get<int> ("max levels", Hierarchy::GetDefaultMaxLevels());
this->graphOutputLevel_ = paramList.get<int> ("debug: graph level", -1);
blockSize_ = paramList.get<int> ("number of equations", 1);
// Save level data
if (paramList.isSublist("print")) {
ParameterList printList = paramList.sublist("print");
if (printList.isParameter("A"))
this->matricesToPrint_ = Teuchos::getArrayFromStringParameter<int>(printList, "A");
if (printList.isParameter("P"))
this->prolongatorsToPrint_ = Teuchos::getArrayFromStringParameter<int>(printList, "P");
if (printList.isParameter("R"))
this->restrictorsToPrint_ = Teuchos::getArrayFromStringParameter<int>(printList, "R");
}
// Translate verbosity parameter
this->verbosity_ = static_cast<MsgType>(Hierarchy::GetDefaultVerbLevel()); // cast int to enum
if (paramList.isParameter("verbosity")) {
std::map<std::string,MsgType> verbMap;
verbMap["none"] = None;
verbMap["low"] = Low;
verbMap["medium"] = Medium;
verbMap["high"] = High;
verbMap["extreme"] = Extreme;
verbMap["test"] = Test;
std::string verbosityLevel = paramList.get<std::string>("verbosity");
TEUCHOS_TEST_FOR_EXCEPTION(verbMap.count(verbosityLevel) == 0, Exceptions::RuntimeError, "Invalid verbosity level: \"" << verbosityLevel << "\"");
this->verbosity_ = verbMap[verbosityLevel];
this->SetVerbLevel(this->verbosity_);
}
// Detect if we need to transfer coordinates to coarse levels. We do that iff
// - we use "laplacian" dropping on some level, or
// - we use repartitioning on some level
// This is not ideal, as we may have "repartition: enable" turned on by default
// and not present in the list, but it is better than nothing.
useCoordinates_ = false;
if ((paramList.isParameter("repartition: enable") && paramList.get<bool>("repartition: enable") == true) ||
(paramList.isParameter("aggregation: drop scheme") && paramList.get<std::string>("aggregation: drop scheme") == "laplacian")) {
useCoordinates_ = true;
} else {
for (int levelID = 0; levelID < this->numDesiredLevel_; levelID++) {
std::string levelStr = "level" + toString(levelID);
if (paramList.isSublist(levelStr)) {
const ParameterList& levelList = paramList.sublist(levelStr);
if ((levelList.isParameter("repartition: enable") && levelList.get<bool>("repartition: enable") == true) ||
(levelList.isParameter("aggregation: drop scheme") && levelList.get<std::string>("aggregation: drop scheme") == "laplacian")) {
useCoordinates_ = true;
break;
}
}
}
}
// Detect if we do implicit P and R rebalance
if (paramList.isParameter("repartition: enable") && paramList.get<bool>("repartition: enable") == true)
this->doPRrebalance_ = paramList.get<bool>("repartition: rebalance P and R", Hierarchy::GetDefaultPRrebalance());
this->implicitTranspose_ = paramList.get<bool>("transpose: use implicit", Hierarchy::GetDefaultImplicitTranspose());
// Create default manager
RCP<FactoryManager> defaultManager = rcp(new FactoryManager());
defaultManager->SetVerbLevel(this->verbosity_);
UpdateFactoryManager(paramList, ParameterList(), *defaultManager);
defaultManager->Print();
for (int levelID = 0; levelID < this->numDesiredLevel_; levelID++) {
RCP<FactoryManager> levelManager;
if (paramList.isSublist("level " + toString(levelID))) {
// Some level specific parameters, update default manager
bool mustAlreadyExist = true;
ParameterList& levelList = paramList.sublist("level " + toString(levelID), mustAlreadyExist);
levelManager = rcp(new FactoryManager(*defaultManager));
levelManager->SetVerbLevel(defaultManager->GetVerbLevel());
UpdateFactoryManager(levelList, paramList, *levelManager);
} else {
// No level specific parameter, use default manager
levelManager = defaultManager;
}
this->AddFactoryManager(levelID, 1, levelManager);
}
if (paramList.isParameter("strict parameter checking") &&
paramList.get<bool> ("strict parameter checking")) {
ParameterList unusedParamList;
// Check for unused parameters that aren't lists
for (ParameterList::ConstIterator itr = paramList.begin(); itr != paramList.end(); ++itr) {
const ParameterEntry& entry = paramList.entry(itr);
if (!entry.isList() && !entry.isUsed())
unusedParamList.setEntry(paramList.name(itr), entry);
}
#if 0
// Check for unused parameters in level-specific sublists
for (int levelID = 0; levelID < this->numDesiredLevel_; levelID++) {
std::string levelStr = "level" + toString(levelID);
if (paramList.isSublist(levelStr)) {
const ParameterList& levelList = paramList.sublist(levelStr);
for (ParameterList::ConstIterator itr = levelList.begin(); itr != levelList.end(); ++itr) {
const ParameterEntry& entry = levelList.entry(itr);
if (!entry.isList() && !entry.isUsed())
unusedParamList.sublist(levelStr).setEntry(levelList.name(itr), entry);
}
}
}
#endif
if (unusedParamList.numParams() > 0) {
std::ostringstream unusedParamsStream;
int indent = 4;
unusedParamList.print(unusedParamsStream, indent);
TEUCHOS_TEST_FOR_EXCEPTION_PURE_MSG(true, Teuchos::Exceptions::InvalidParameter,
"WARNING: Unused parameters were detected. Please check spelling and type." << std::endl << unusedParamsStream.str());
}
}
// FIXME: parameters passed to packages, like Ifpack2, are not touched by us, resulting in "[unused]" flag
// being displayed. On the other hand, we don't want to simply iterate through them touching. I don't know
// what a good solution looks like
this->GetOStream(static_cast<MsgType>(Runtime1 | Test), 0) << paramList << std::endl;
}
void HierarchyUtils<Scalar, LocalOrdinal, GlobalOrdinal, Node>::AddNonSerializableDataToHierarchy(HierarchyManager& HM, Hierarchy& H, const ParameterList& paramList) {
for (ParameterList::ConstIterator it = paramList.begin(); it != paramList.end(); it++) {
const std::string& levelName = it->first;
// Check for mach of the form "level X" where X is a positive integer
if (paramList.isSublist(levelName) && levelName.find("level ") == 0 && levelName.size() > 6) {
int levelID = strtol(levelName.substr(6).c_str(), 0, 0);
if (levelID > 0)
{
// Do enough level adding so we can be sure to add the data to the right place
for (int i = H.GetNumLevels(); i <= levelID; i++)
H.AddNewLevel();
}
RCP<Level> level = H.GetLevel(levelID);
RCP<FactoryManager> M = Teuchos::rcp_dynamic_cast<FactoryManager>(HM.GetFactoryManager(levelID));
TEUCHOS_TEST_FOR_EXCEPTION(M.is_null(), Exceptions::InvalidArgument, "MueLu::Utils::AddNonSerializableDataToHierarchy: cannot get FactoryManager");
// Grab the level sublist & loop over parameters
const ParameterList& levelList = paramList.sublist(levelName);
for (ParameterList::ConstIterator it2 = levelList.begin(); it2 != levelList.end(); it2++) {
const std::string& name = it2->first;
TEUCHOS_TEST_FOR_EXCEPTION(name != "A" && name != "P" && name != "R" &&
name != "Nullspace" && name != "Coordinates" &&
!IsParamMuemexVariable(name), Exceptions::InvalidArgument,
"MueLu::Utils::AddNonSerializableDataToHierarchy: parameter list contains unknown data type");
if (name == "A") {
level->Set(name, Teuchos::getValue<RCP<Matrix > > (it2->second),NoFactory::get());
M->SetFactory(name, NoFactory::getRCP()); // TAW: not sure about this: be aware that this affects all levels
// However, A is accessible through NoFactory anyway, so it should
// be fine here.
}
else if( name == "P" || name == "R") {
level->AddKeepFlag(name,NoFactory::get(),MueLu::UserData);
level->Set(name, Teuchos::getValue<RCP<Matrix > > (it2->second), M->GetFactory(name).get());
}
else if (name == "Nullspace")
{
level->AddKeepFlag(name,NoFactory::get(),MueLu::UserData);
level->Set(name, Teuchos::getValue<RCP<MultiVector > >(it2->second), NoFactory::get());
//M->SetFactory(name, NoFactory::getRCP()); // TAW: generally it is a bad idea to overwrite the factory manager data here
// One should do this only in very special cases
}
else if(name == "Coordinates") //Scalar of Coordinates MV is always double
{
level->AddKeepFlag(name,NoFactory::get(),MueLu::UserData);
level->Set(name, Teuchos::getValue<RCP<Xpetra::MultiVector<double, LocalOrdinal, GlobalOrdinal, Node> > >(it2->second), NoFactory::get());
//M->SetFactory(name, NoFactory::getRCP()); // TAW: generally it is a bad idea to overwrite the factory manager data here
}
#ifdef HAVE_MUELU_MATLAB
else
{
//Custom variable for Muemex
size_t typeNameStart = name.find_first_not_of(' ');
size_t typeNameEnd = name.find(' ', typeNameStart);
std::string typeName = name.substr(typeNameStart, typeNameEnd - typeNameStart);
std::transform(typeName.begin(), typeName.end(), typeName.begin(), ::tolower);
level->AddKeepFlag(name, NoFactory::get(), MueLu::UserData);
if(typeName == "matrix")
level->Set(name, Teuchos::getValue<RCP<Matrix> >(it2->second), NoFactory::get());
else if(typeName == "multivector")
level->Set(name, Teuchos::getValue<RCP<MultiVector> >(it2->second), NoFactory::get());
else if(typeName == "map")
level->Set(name, Teuchos::getValue<RCP<Xpetra::Map<LocalOrdinal, GlobalOrdinal, Node> > >(it2->second), NoFactory::get());
else if(typeName == "ordinalvector")
level->Set(name, Teuchos::getValue<RCP<Xpetra::Vector<LocalOrdinal, LocalOrdinal, GlobalOrdinal, Node> > >(it2->second), NoFactory::get());
else if(typeName == "scalar")
level->Set(name, Teuchos::getValue<Scalar>(it2->second), NoFactory::get());
else if(typeName == "double")
level->Set(name, Teuchos::getValue<double>(it2->second), NoFactory::get());
else if(typeName == "complex")
level->Set(name, Teuchos::getValue<std::complex<double> >(it2->second), NoFactory::get());
else if(typeName == "int")
level->Set(name, Teuchos::getValue<int>(it2->second), NoFactory::get());
else if(typeName == "string")
level->Set(name, Teuchos::getValue<std::string>(it2->second), NoFactory::get());
}
#endif
}
}
}
}
Teuchos::RCP<MueLu::TpetraOperator<Scalar,LocalOrdinal,GlobalOrdinal,Node> >
CreateTpetraPreconditioner(const Teuchos::RCP<Tpetra::CrsMatrix <Scalar, LocalOrdinal, GlobalOrdinal, Node> >& inA,
Teuchos::ParameterList& paramListIn,
const Teuchos::RCP<Tpetra::MultiVector<double, LocalOrdinal, GlobalOrdinal, Node> >& inCoords = Teuchos::null,
const Teuchos::RCP<Tpetra::MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node> >& inNullspace = Teuchos::null)
{
typedef Scalar SC;
typedef LocalOrdinal LO;
typedef GlobalOrdinal GO;
typedef Node NO;
using Teuchos::ParameterList;
typedef Xpetra::MultiVector<SC,LO,GO,NO> MultiVector;
typedef Xpetra::Matrix<SC,LO,GO,NO> Matrix;
typedef Hierarchy<SC,LO,GO,NO> Hierarchy;
typedef HierarchyManager<SC,LO,GO,NO> HierarchyManager;
bool hasParamList = paramListIn.numParams();
RCP<HierarchyManager> mueLuFactory;
ParameterList paramList = paramListIn;
std::string syntaxStr = "parameterlist: syntax";
if (hasParamList && paramList.isParameter(syntaxStr) && paramList.get<std::string>(syntaxStr) == "ml") {
paramList.remove(syntaxStr);
mueLuFactory = rcp(new MLParameterListInterpreter<SC,LO,GO,NO>(paramList));
} else {
mueLuFactory = rcp(new ParameterListInterpreter <SC,LO,GO,NO>(paramList));
}
RCP<Hierarchy> H = mueLuFactory->CreateHierarchy();
H->setlib(Xpetra::UseTpetra);
// Wrap A
RCP<Matrix> A = TpetraCrs_To_XpetraMatrix<SC,LO,GO,NO>(inA);
H->GetLevel(0)->Set("A", A);
// Wrap coordinates if available
if (inCoords != Teuchos::null) {
RCP<Xpetra::MultiVector<double,LO,GO,NO> > coordinates = TpetraMultiVector_To_XpetraMultiVector<double,LO,GO,NO>(inCoords);
H->GetLevel(0)->Set("Coordinates", coordinates);
}
// Wrap nullspace if available, otherwise use constants
RCP<MultiVector> nullspace;
if (inNullspace != Teuchos::null) {
nullspace = TpetraMultiVector_To_XpetraMultiVector<SC,LO,GO,NO>(inNullspace);
} else {
int nPDE = MasterList::getDefault<int>("number of equations");
if (paramList.isSublist("Matrix")) {
// Factory style parameter list
const Teuchos::ParameterList& operatorList = paramList.sublist("Matrix");
if (operatorList.isParameter("PDE equations"))
nPDE = operatorList.get<int>("PDE equations");
} else if (paramList.isParameter("number of equations")) {
// Easy style parameter list
nPDE = paramList.get<int>("number of equations");
}
nullspace = Xpetra::MultiVectorFactory<SC,LO,GO,NO>::Build(A->getDomainMap(), nPDE);
if (nPDE == 1) {
nullspace->putScalar(Teuchos::ScalarTraits<SC>::one());
} else {
for (int i = 0; i < nPDE; i++) {
Teuchos::ArrayRCP<SC> nsData = nullspace->getDataNonConst(i);
for (int j = 0; j < nsData.size(); j++) {
GO GID = A->getDomainMap()->getGlobalElement(j) - A->getDomainMap()->getIndexBase();
if ((GID-i) % nPDE == 0)
nsData[j] = Teuchos::ScalarTraits<SC>::one();
}
}
}
}
H->GetLevel(0)->Set("Nullspace", nullspace);
Teuchos::ParameterList nonSerialList,dummyList;
ExtractNonSerializableData(paramList, dummyList, nonSerialList);
HierarchyUtils<SC,LO,GO,NO>::AddNonSerializableDataToHierarchy(*mueLuFactory,*H, nonSerialList);
mueLuFactory->SetupHierarchy(*H);
return rcp(new TpetraOperator<SC,LO,GO,NO>(H));
}
Teuchos::RCP<MueLu::TpetraOperator<Scalar,LocalOrdinal,GlobalOrdinal,Node> >
CreateTpetraPreconditioner(const Teuchos::RCP<Tpetra::Operator<Scalar, LocalOrdinal, GlobalOrdinal, Node> > &inA,
Teuchos::ParameterList& inParamList,
const Teuchos::RCP<Tpetra::MultiVector<double, LocalOrdinal, GlobalOrdinal, Node>>& inCoords = Teuchos::null,
const Teuchos::RCP<Tpetra::MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>>& inNullspace = Teuchos::null)
{
typedef Scalar SC;
typedef LocalOrdinal LO;
typedef GlobalOrdinal GO;
typedef Node NO;
using Teuchos::ParameterList;
typedef Xpetra::MultiVector<SC,LO,GO,NO> MultiVector;
typedef Xpetra::Matrix<SC,LO,GO,NO> Matrix;
typedef Hierarchy<SC,LO,GO,NO> Hierarchy;
typedef HierarchyManager<SC,LO,GO,NO> HierarchyManager;
typedef Tpetra::CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> crs_matrix_type;
typedef Tpetra::Experimental::BlockCrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> block_crs_matrix_type;
bool hasParamList = inParamList.numParams();
RCP<HierarchyManager> mueLuFactory;
ParameterList paramList = inParamList;
RCP<const crs_matrix_type> constCrsA;
RCP<crs_matrix_type> crsA;
#if defined(HAVE_MUELU_EXPERIMENTAL) and defined(HAVE_MUELU_AMGX)
std::string externalMG = "use external multigrid package";
if (hasParamList && paramList.isParameter(externalMG) && paramList.get<std::string>(externalMG) == "amgx"){
constCrsA = rcp_dynamic_cast<const crs_matrix_type>(inA);
TEUCHOS_TEST_FOR_EXCEPTION(constCrsA == Teuchos::null, Exceptions::RuntimeError, "CreateTpetraPreconditioner: failed to dynamic cast to Tpetra::CrsMatrix, which is required to be able to use AmgX.");
return rcp(new AMGXOperator<SC,LO,GO,NO>(inA,inParamList));
}
#endif
std::string syntaxStr = "parameterlist: syntax";
if (hasParamList && paramList.isParameter(syntaxStr) && paramList.get<std::string>(syntaxStr) == "ml") {
paramList.remove(syntaxStr);
mueLuFactory = rcp(new MLParameterListInterpreter<SC,LO,GO,NO>(paramList));
} else {
mueLuFactory = rcp(new ParameterListInterpreter <SC,LO,GO,NO>(paramList,inA->getDomainMap()->getComm()));
}
RCP<Hierarchy> H = mueLuFactory->CreateHierarchy();
H->setlib(Xpetra::UseTpetra);
// Wrap A
RCP<Matrix> A;
RCP<block_crs_matrix_type> bcrsA = rcp_dynamic_cast<block_crs_matrix_type>(inA);
crsA = rcp_dynamic_cast<crs_matrix_type>(inA);
if (crsA != Teuchos::null)
A = TpetraCrs_To_XpetraMatrix<SC,LO,GO,NO>(crsA);
else if (bcrsA != Teuchos::null) {
RCP<Xpetra::CrsMatrix<SC,LO,GO,NO> > temp = rcp(new Xpetra::TpetraBlockCrsMatrix<SC,LO,GO,NO>(bcrsA));
TEUCHOS_TEST_FOR_EXCEPTION(temp==Teuchos::null, Exceptions::RuntimeError, "CreateTpetraPreconditioner: cast from Tpetra::Experimental::BlockCrsMatrix to Xpetra::TpetraBlockCrsMatrix failed.");
A = rcp(new Xpetra::CrsMatrixWrap<SC,LO,GO,NO>(temp));
}
else {
TEUCHOS_TEST_FOR_EXCEPTION(true, Exceptions::RuntimeError, "CreateTpetraPreconditioner: only Tpetra CrsMatrix and BlockCrsMatrix types are supported.");
}
H->GetLevel(0)->Set("A", A);
// Wrap coordinates if available
if (inCoords != Teuchos::null) {
RCP<Xpetra::MultiVector<double,LO,GO,NO> > coordinates = TpetraMultiVector_To_XpetraMultiVector<double,LO,GO,NO>(inCoords);
H->GetLevel(0)->Set("Coordinates", coordinates);
}
// Wrap nullspace if available, otherwise use constants
RCP<MultiVector> nullspace;
if (inNullspace != Teuchos::null) {
nullspace = TpetraMultiVector_To_XpetraMultiVector<SC,LO,GO,NO>(inNullspace);
} else {
int nPDE = MasterList::getDefault<int>("number of equations");
if (paramList.isSublist("Matrix")) {
// Factory style parameter list
const Teuchos::ParameterList& operatorList = paramList.sublist("Matrix");
if (operatorList.isParameter("PDE equations"))
nPDE = operatorList.get<int>("PDE equations");
} else if (paramList.isParameter("number of equations")) {
// Easy style parameter list
nPDE = paramList.get<int>("number of equations");
}
nullspace = Xpetra::MultiVectorFactory<SC,LO,GO,NO>::Build(A->getDomainMap(), nPDE);
if (nPDE == 1) {
nullspace->putScalar(Teuchos::ScalarTraits<SC>::one());
} else {
for (int i = 0; i < nPDE; i++) {
Teuchos::ArrayRCP<SC> nsData = nullspace->getDataNonConst(i);
for (int j = 0; j < nsData.size(); j++) {
GO GID = A->getDomainMap()->getGlobalElement(j) - A->getDomainMap()->getIndexBase();
if ((GID-i) % nPDE == 0)
nsData[j] = Teuchos::ScalarTraits<SC>::one();
}
}
}
}
H->GetLevel(0)->Set("Nullspace", nullspace);
Teuchos::ParameterList nonSerialList,dummyList;
ExtractNonSerializableData(paramList, dummyList, nonSerialList);
HierarchyUtils<SC,LO,GO,NO>::AddNonSerializableDataToHierarchy(*mueLuFactory,*H, nonSerialList);
mueLuFactory->SetupHierarchy(*H);
return rcp(new TpetraOperator<SC,LO,GO,NO>(H));
}
int main(int argc, char *argv[]) {
#include <MueLu_UseShortNames.hpp>
using Teuchos::RCP; // reference count pointers
using Teuchos::rcp;
using Teuchos::TimeMonitor;
using Teuchos::ParameterList;
// =========================================================================
// MPI initialization using Teuchos
// =========================================================================
Teuchos::GlobalMPISession mpiSession(&argc, &argv, NULL);
RCP< const Teuchos::Comm<int> > comm = Teuchos::DefaultComm<int>::getComm();
// =========================================================================
// Convenient definitions
// =========================================================================
typedef Teuchos::ScalarTraits<SC> STS;
SC zero = STS::zero(), one = STS::one();
// =========================================================================
// Parameters initialization
// =========================================================================
Teuchos::CommandLineProcessor clp(false);
GO nx = 100, ny = 100, nz = 100;
Galeri::Xpetra::Parameters<GO> galeriParameters(clp, nx, ny, nz, "Laplace2D"); // manage parameters of the test case
Xpetra::Parameters xpetraParameters(clp); // manage parameters of Xpetra
std::string xmlFileName = "scalingTest.xml"; clp.setOption("xml", &xmlFileName, "read parameters from a file [default = 'scalingTest.xml']");
bool printTimings = true; clp.setOption("timings", "notimings", &printTimings, "print timings to screen");
int writeMatricesOPT = -2; clp.setOption("write", &writeMatricesOPT, "write matrices to file (-1 means all; i>=0 means level i)");
std::string dsolveType = "cg", solveType; clp.setOption("solver", &dsolveType, "solve type: (none | cg | gmres | standalone)");
double dtol = 1e-12, tol; clp.setOption("tol", &dtol, "solver convergence tolerance");
std::string mapFile; clp.setOption("map", &mapFile, "map data file");
std::string matrixFile; clp.setOption("matrix", &matrixFile, "matrix data file");
std::string coordFile; clp.setOption("coords", &coordFile, "coordinates data file");
int numRebuilds = 0; clp.setOption("rebuild", &numRebuilds, "#times to rebuild hierarchy");
int maxIts = 200; clp.setOption("its", &maxIts, "maximum number of solver iterations");
bool scaleResidualHistory = true; clp.setOption("scale", "noscale", &scaleResidualHistory, "scaled Krylov residual history");
switch (clp.parse(argc, argv)) {
case Teuchos::CommandLineProcessor::PARSE_HELP_PRINTED: return EXIT_SUCCESS;
case Teuchos::CommandLineProcessor::PARSE_ERROR:
case Teuchos::CommandLineProcessor::PARSE_UNRECOGNIZED_OPTION: return EXIT_FAILURE;
case Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL: break;
}
Xpetra::UnderlyingLib lib = xpetraParameters.GetLib();
ParameterList paramList;
Teuchos::updateParametersFromXmlFileAndBroadcast(xmlFileName, Teuchos::Ptr<ParameterList>(¶mList), *comm);
bool isDriver = paramList.isSublist("Run1");
if (isDriver) {
// update galeriParameters with the values from the XML file
ParameterList& realParams = galeriParameters.GetParameterList();
for (ParameterList::ConstIterator it = realParams.begin(); it != realParams.end(); it++) {
const std::string& name = realParams.name(it);
if (paramList.isParameter(name))
realParams.setEntry(name, paramList.getEntry(name));
}
}
// Retrieve matrix parameters (they may have been changed on the command line)
// [for instance, if we changed matrix type from 2D to 3D we need to update nz]
ParameterList galeriList = galeriParameters.GetParameterList();
// =========================================================================
// Problem construction
// =========================================================================
std::ostringstream galeriStream;
comm->barrier();
RCP<TimeMonitor> globalTimeMonitor = rcp(new TimeMonitor(*TimeMonitor::getNewTimer("ScalingTest: S - Global Time")));
RCP<TimeMonitor> tm = rcp(new TimeMonitor(*TimeMonitor::getNewTimer("ScalingTest: 1 - Matrix Build")));
RCP<Matrix> A;
RCP<const Map> map;
RCP<MultiVector> coordinates;
RCP<MultiVector> nullspace;
if (matrixFile.empty()) {
galeriStream << "========================================================\n" << xpetraParameters << galeriParameters;
// Galeri will attempt to create a square-as-possible distribution of subdomains di, e.g.,
// d1 d2 d3
// d4 d5 d6
// d7 d8 d9
// d10 d11 d12
// A perfect distribution is only possible when the #processors is a perfect square.
// This *will* result in "strip" distribution if the #processors is a prime number or if the factors are very different in
// size. For example, np=14 will give a 7-by-2 distribution.
// If you don't want Galeri to do this, specify mx or my on the galeriList.
std::string matrixType = galeriParameters.GetMatrixType();
// Create map and coordinates
// In the future, we hope to be able to first create a Galeri problem, and then request map and coordinates from it
// At the moment, however, things are fragile as we hope that the Problem uses same map and coordinates inside
if (matrixType == "Laplace1D") {
map = Galeri::Xpetra::CreateMap<LO, GO, Node>(xpetraParameters.GetLib(), "Cartesian1D", comm, galeriList);
coordinates = Galeri::Xpetra::Utils::CreateCartesianCoordinates<SC,LO,GO,Map,MultiVector>("1D", map, galeriList);
} else if (matrixType == "Laplace2D" || matrixType == "Star2D" ||
matrixType == "BigStar2D" || matrixType == "Elasticity2D") {
map = Galeri::Xpetra::CreateMap<LO, GO, Node>(xpetraParameters.GetLib(), "Cartesian2D", comm, galeriList);
coordinates = Galeri::Xpetra::Utils::CreateCartesianCoordinates<SC,LO,GO,Map,MultiVector>("2D", map, galeriList);
} else if (matrixType == "Laplace3D" || matrixType == "Brick3D" || matrixType == "Elasticity3D") {
map = Galeri::Xpetra::CreateMap<LO, GO, Node>(xpetraParameters.GetLib(), "Cartesian3D", comm, galeriList);
coordinates = Galeri::Xpetra::Utils::CreateCartesianCoordinates<SC,LO,GO,Map,MultiVector>("3D", map, galeriList);
}
// Expand map to do multiple DOF per node for block problems
if (matrixType == "Elasticity2D")
map = Xpetra::MapFactory<LO,GO,Node>::Build(map, 2);
if (matrixType == "Elasticity3D")
map = Xpetra::MapFactory<LO,GO,Node>::Build(map, 3);
galeriStream << "Processor subdomains in x direction: " << galeriList.get<int>("mx") << std::endl
<< "Processor subdomains in y direction: " << galeriList.get<int>("my") << std::endl
<< "Processor subdomains in z direction: " << galeriList.get<int>("mz") << std::endl
<< "========================================================" << std::endl;
if (matrixType == "Elasticity2D" || matrixType == "Elasticity3D") {
// Our default test case for elasticity: all boundaries of a square/cube have Neumann b.c. except left which has Dirichlet
galeriList.set("right boundary" , "Neumann");
galeriList.set("bottom boundary", "Neumann");
galeriList.set("top boundary" , "Neumann");
galeriList.set("front boundary" , "Neumann");
galeriList.set("back boundary" , "Neumann");
}
RCP<Galeri::Xpetra::Problem<Map,CrsMatrixWrap,MultiVector> > Pr =
Galeri::Xpetra::BuildProblem<SC,LO,GO,Map,CrsMatrixWrap,MultiVector>(galeriParameters.GetMatrixType(), map, galeriList);
A = Pr->BuildMatrix();
nullspace = MultiVectorFactory::Build(map, 1);
if (matrixType == "Elasticity2D" ||
matrixType == "Elasticity3D") {
nullspace = Pr->BuildNullspace();
A->SetFixedBlockSize((galeriParameters.GetMatrixType() == "Elasticity2D") ? 2 : 3);
} else {
nullspace->putScalar(one);
}
} else {
if (!mapFile.empty())
map = Utils2::ReadMap(mapFile, xpetraParameters.GetLib(), comm);
comm->barrier();
if (lib == Xpetra::UseEpetra) {
A = Utils::Read(matrixFile, map);
} else {
// Tpetra matrix reader is still broken, so instead we read in
// a matrix in a binary format and then redistribute it
const bool binaryFormat = true;
A = Utils::Read(matrixFile, lib, comm, binaryFormat);
RCP<Matrix> newMatrix = MatrixFactory::Build(map, 1);
RCP<Import> importer = ImportFactory::Build(A->getRowMap(), map);
newMatrix->doImport(*A, *importer, Xpetra::INSERT);
newMatrix->fillComplete();
A.swap(newMatrix);
}
comm->barrier();
if (!coordFile.empty())
coordinates = Utils2::ReadMultiVector(coordFile, map);
nullspace = MultiVectorFactory::Build(map, 1);
nullspace->putScalar(one);
}
comm->barrier();
tm = Teuchos::null;
galeriStream << "Galeri complete.\n========================================================" << std::endl;
int numReruns = 1;
if (paramList.isParameter("number of reruns"))
numReruns = paramList.get<int>("number of reruns");
const bool mustAlreadyExist = true;
for (int rerunCount = 1; rerunCount <= numReruns; rerunCount++) {
ParameterList mueluList, runList;
bool stop = false;
if (isDriver) {
runList = paramList.sublist("Run1", mustAlreadyExist);
mueluList = runList .sublist("MueLu", mustAlreadyExist);
} else {
mueluList = paramList;
stop = true;
}
int runCount = 1;
do {
A->SetMaxEigenvalueEstimate(-one);
solveType = dsolveType;
tol = dtol;
int savedOut = -1;
FILE* openedOut = NULL;
if (isDriver) {
if (runList.isParameter("filename")) {
// Redirect all output into a filename We have to redirect all output,
// including printf's, therefore we cannot simply replace C++ cout
// buffers, and have to use heavy machinary (dup2)
std::string filename = runList.get<std::string>("filename");
if (numReruns > 1)
filename += "_run" + MueLu::toString(rerunCount);
filename += (lib == Xpetra::UseEpetra ? ".epetra" : ".tpetra");
savedOut = dup(STDOUT_FILENO);
openedOut = fopen(filename.c_str(), "w");
dup2(fileno(openedOut), STDOUT_FILENO);
}
if (runList.isParameter("solver")) solveType = runList.get<std::string>("solver");
if (runList.isParameter("tol")) tol = runList.get<double> ("tol");
}
// Instead of checking each time for rank, create a rank 0 stream
RCP<Teuchos::FancyOStream> fancy = Teuchos::fancyOStream(Teuchos::rcpFromRef(std::cout));
Teuchos::FancyOStream& fancyout = *fancy;
fancyout.setOutputToRootOnly(0);
fancyout << galeriStream.str();
// =========================================================================
// Preconditioner construction
// =========================================================================
comm->barrier();
tm = rcp(new TimeMonitor(*TimeMonitor::getNewTimer("ScalingTest: 1.5 - MueLu read XML")));
RCP<HierarchyManager> mueLuFactory = rcp(new ParameterListInterpreter(mueluList));
comm->barrier();
tm = rcp(new TimeMonitor(*TimeMonitor::getNewTimer("ScalingTest: 2 - MueLu Setup")));
RCP<Hierarchy> H;
for (int i = 0; i <= numRebuilds; i++) {
A->SetMaxEigenvalueEstimate(-one);
H = mueLuFactory->CreateHierarchy();
H->GetLevel(0)->Set("A", A);
H->GetLevel(0)->Set("Nullspace", nullspace);
if (!coordinates.is_null())
H->GetLevel(0)->Set("Coordinates", coordinates);
mueLuFactory->SetupHierarchy(*H);
}
comm->barrier();
tm = Teuchos::null;
// =========================================================================
// System solution (Ax = b)
// =========================================================================
comm->barrier();
tm = rcp(new TimeMonitor(*TimeMonitor::getNewTimer("ScalingTest: 3 - LHS and RHS initialization")));
RCP<Vector> X = VectorFactory::Build(map);
RCP<Vector> B = VectorFactory::Build(map);
{
// we set seed for reproducibility
Utils::SetRandomSeed(*comm);
X->randomize();
A->apply(*X, *B, Teuchos::NO_TRANS, one, zero);
Teuchos::Array<STS::magnitudeType> norms(1);
B->norm2(norms);
B->scale(one/norms[0]);
X->putScalar(zero);
}
tm = Teuchos::null;
if (writeMatricesOPT > -2) {
tm = rcp(new TimeMonitor(*TimeMonitor::getNewTimer("ScalingTest: 3.5 - Matrix output")));
H->Write(writeMatricesOPT, writeMatricesOPT);
tm = Teuchos::null;
}
comm->barrier();
if (solveType == "none") {
// Do not perform a solve
} else if (solveType == "standalone") {
tm = rcp (new TimeMonitor(*TimeMonitor::getNewTimer("ScalingTest: 4 - Fixed Point Solve")));
H->IsPreconditioner(false);
H->Iterate(*B, *X, maxIts);
} else if (solveType == "cg" || solveType == "gmres") {
#ifdef HAVE_MUELU_BELOS
tm = rcp(new TimeMonitor(*TimeMonitor::getNewTimer("ScalingTest: 5 - Belos Solve")));
// Operator and Multivector type that will be used with Belos
typedef MultiVector MV;
typedef Belos::OperatorT<MV> OP;
H->IsPreconditioner(true);
// Define Operator and Preconditioner
Teuchos::RCP<OP> belosOp = Teuchos::rcp(new Belos::XpetraOp<SC, LO, GO, NO, LMO>(A)); // Turns a Xpetra::Matrix object into a Belos operator
Teuchos::RCP<OP> belosPrec = Teuchos::rcp(new Belos::MueLuOp <SC, LO, GO, NO, LMO>(H)); // Turns a MueLu::Hierarchy object into a Belos operator
// Construct a Belos LinearProblem object
RCP< Belos::LinearProblem<SC, MV, OP> > belosProblem = rcp(new Belos::LinearProblem<SC, MV, OP>(belosOp, X, B));
belosProblem->setRightPrec(belosPrec);
bool set = belosProblem->setProblem();
if (set == false) {
fancyout << "\nERROR: Belos::LinearProblem failed to set up correctly!" << std::endl;
return EXIT_FAILURE;
}
// Belos parameter list
Teuchos::ParameterList belosList;
belosList.set("Maximum Iterations", maxIts); // Maximum number of iterations allowed
belosList.set("Convergence Tolerance", tol); // Relative convergence tolerance requested
belosList.set("Verbosity", Belos::Errors + Belos::Warnings + Belos::StatusTestDetails);
belosList.set("Output Frequency", 1);
belosList.set("Output Style", Belos::Brief);
if (!scaleResidualHistory)
belosList.set("Implicit Residual Scaling", "None");
// Create an iterative solver manager
RCP< Belos::SolverManager<SC, MV, OP> > solver;
if (solveType == "cg") {
solver = rcp(new Belos::PseudoBlockCGSolMgr <SC, MV, OP>(belosProblem, rcp(&belosList, false)));
} else if (solveType == "gmres") {
solver = rcp(new Belos::BlockGmresSolMgr<SC, MV, OP>(belosProblem, rcp(&belosList, false)));
}
// Perform solve
Belos::ReturnType ret = Belos::Unconverged;
try {
ret = solver->solve();
// Get the number of iterations for this solve.
fancyout << "Number of iterations performed for this solve: " << solver->getNumIters() << std::endl;
} catch(...) {
fancyout << std::endl << "ERROR: Belos threw an error! " << std::endl;
}
// Check convergence
if (ret != Belos::Converged)
fancyout << std::endl << "ERROR: Belos did not converge! " << std::endl;
else
fancyout << std::endl << "SUCCESS: Belos converged!" << std::endl;
#endif //ifdef HAVE_MUELU_BELOS
} else {
throw MueLu::Exceptions::RuntimeError("Unknown solver type: \"" + solveType + "\"");
}
comm->barrier();
tm = Teuchos::null;
globalTimeMonitor = Teuchos::null;
if (printTimings)
TimeMonitor::summarize(A->getRowMap()->getComm().ptr(), std::cout, false, true, false, Teuchos::Union);
TimeMonitor::clearCounters();
if (isDriver) {
if (openedOut != NULL) {
dup2(savedOut, STDOUT_FILENO);
fclose(openedOut);
openedOut = NULL;
}
try {
runList = paramList.sublist("Run" + MueLu::toString(++runCount), mustAlreadyExist);
mueluList = runList .sublist("MueLu", mustAlreadyExist);
} catch (std::exception) {
stop = true;
}
}
} while (stop == false);
}
return 0;
} //main
RCP<LinearOpWithSolveFactoryBase<double> >
LOWSFactoryBuilder::createLOWSFactory(const ParameterList& params)
{
/* check that we have a linear solver parameter list */
// TEST_FOR_EXCEPTION(params.name() != "Linear Solver",
// std::runtime_error,
// "Expected \"Linear Solver\" as name of parameter list input "
// "to createLOWSFactory()");
RCP<LinearOpWithSolveFactoryBase<double> > rtn;
RCP<PreconditionerFactoryBase<double> > prec;
if (params.isSublist("Amesos"))
{
RCP<ParameterList> p = rcp(new ParameterList(params.sublist("Amesos")));
rtn = rcp(new AmesosLinearOpWithSolveFactory());
rtn->setParameterList(p);
}
else if (params.isSublist("Aztec"))
{
RCP<ParameterList> p = rcp(new ParameterList(params.sublist("Aztec")));
rtn = rcp(new AztecOOLinearOpWithSolveFactory());
rtn->setParameterList(p);
}
else if (params.isSublist("Belos"))
{
RCP<ParameterList> p = rcp(new ParameterList(params.sublist("Belos")));
rtn = rcp(new BelosLinearOpWithSolveFactory<double>());
rtn->setParameterList(p);
}
else
{
TEST_FOR_EXCEPTION(true, std::runtime_error,
"solver parameter list did not contain one of [Aztec, Amesos, "
"Belos]");
}
if (params.isSublist("Preconditioner"))
{
ParameterList precParams = params.sublist("Preconditioner");
std::string precType = precParams.get<string>("Type");
if (precType=="ML")
{
std::string probType = getParameter<string>(precParams, "Problem Type");
ParameterList mlParams = precParams.sublist("ML Settings");
prec = rcp(new MLPreconditionerFactory(probType, mlParams));
}
else if (precType=="Ifpack")
{
std::string probType = getParameter<string>(precParams, "Prec Type");
RCP<ParameterList> ifpackParams
= rcp(new ParameterList(precParams.sublist("Ifpack")));
prec = rcp(new IfpackPreconditionerFactory());
prec->setParameterList(ifpackParams);
}
else
{
TEST_FOR_EXCEPTION(true, std::runtime_error,
"Preconditioner type [" << precType << "] not recognized");
}
}
TEST_FOR_EXCEPTION(prec.get() != 0 && !rtn->acceptsPreconditionerFactory(),
std::runtime_error,
"Huh? You have provided a preconditioner for a solver that cannot "
"accept a preconditioner!");
if (prec.get() != 0 && rtn->acceptsPreconditionerFactory())
{
rtn->setPreconditionerFactory(prec, "precond");
}
return rtn;
}
int
main (int argc, char *argv[])
{
using namespace TrilinosCouplings; // Yes, this means I'm lazy.
using TpetraIntrepidPoissonExample::exactResidualNorm;
using TpetraIntrepidPoissonExample::makeMatrixAndRightHandSide;
using TpetraIntrepidPoissonExample::solveWithBelos;
using TpetraIntrepidPoissonExample::solveWithBelosGPU;
using IntrepidPoissonExample::makeMeshInput;
using IntrepidPoissonExample::parseCommandLineArguments;
using IntrepidPoissonExample::setCommandLineArgumentDefaults;
using IntrepidPoissonExample::setMaterialTensorOffDiagonalValue;
using IntrepidPoissonExample::setUpCommandLineArguments;
using Tpetra::DefaultPlatform;
using Teuchos::Comm;
using Teuchos::outArg;
using Teuchos::ParameterList;
using Teuchos::parameterList;
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcpFromRef;
using Teuchos::getFancyOStream;
using Teuchos::FancyOStream;
using std::endl;
// Pull in typedefs from the example's namespace.
typedef TpetraIntrepidPoissonExample::ST ST;
#ifdef HAVE_TRILINOSCOUPLINGS_MUELU
typedef TpetraIntrepidPoissonExample::LO LO;
typedef TpetraIntrepidPoissonExample::GO GO;
#endif // HAVE_TRILINOSCOUPLINGS_MUELU
typedef TpetraIntrepidPoissonExample::Node Node;
typedef Teuchos::ScalarTraits<ST> STS;
typedef STS::magnitudeType MT;
typedef Teuchos::ScalarTraits<MT> STM;
typedef TpetraIntrepidPoissonExample::sparse_matrix_type sparse_matrix_type;
typedef TpetraIntrepidPoissonExample::vector_type vector_type;
typedef TpetraIntrepidPoissonExample::operator_type operator_type;
bool success = true;
try {
Teuchos::oblackholestream blackHole;
Teuchos::GlobalMPISession mpiSession (&argc, &argv, &blackHole);
const int myRank = mpiSession.getRank ();
//const int numProcs = mpiSession.getNProc ();
// Get the default communicator and Kokkos Node instance
RCP<const Comm<int> > comm =
DefaultPlatform::getDefaultPlatform ().getComm ();
RCP<Node> node = DefaultPlatform::getDefaultPlatform ().getNode ();
// Did the user specify --help at the command line to print help
// with command-line arguments?
bool printedHelp = false;
// Values of command-line arguments.
int nx, ny, nz;
std::string xmlInputParamsFile;
bool verbose, debug;
int maxNumItersFromCmdLine = -1; // -1 means "read from XML file"
double tolFromCmdLine = -1.0; // -1 means "read from XML file"
std::string solverName = "GMRES";
ST materialTensorOffDiagonalValue = 0.0;
// Set default values of command-line arguments.
setCommandLineArgumentDefaults (nx, ny, nz, xmlInputParamsFile,
solverName, verbose, debug);
// Parse and validate command-line arguments.
Teuchos::CommandLineProcessor cmdp (false, true);
setUpCommandLineArguments (cmdp, nx, ny, nz, xmlInputParamsFile,
solverName, tolFromCmdLine,
maxNumItersFromCmdLine,
verbose, debug);
cmdp.setOption ("materialTensorOffDiagonalValue",
&materialTensorOffDiagonalValue, "Off-diagonal value in "
"the material tensor. This controls the iteration count. "
"Be careful with this if you use CG, since you can easily "
"make the matrix indefinite.");
// Additional command-line arguments for GPU experimentation.
bool gpu = false;
cmdp.setOption ("gpu", "no-gpu", &gpu,
"Run example using GPU node (if supported)");
int ranks_per_node = 1;
cmdp.setOption ("ranks_per_node", &ranks_per_node,
"Number of MPI ranks per node");
int gpu_ranks_per_node = 1;
cmdp.setOption ("gpu_ranks_per_node", &gpu_ranks_per_node,
"Number of MPI ranks per node for GPUs");
int device_offset = 0;
cmdp.setOption ("device_offset", &device_offset,
"Offset for attaching MPI ranks to CUDA devices");
// Additional command-line arguments for dumping the generated
// matrix or its row Map to output files.
//
// FIXME (mfh 09 Apr 2014) Need to port these command-line
// arguments to the Epetra version.
// If matrixFilename is nonempty, dump the matrix to that file
// in MatrixMarket format.
std::string matrixFilename;
cmdp.setOption ("matrixFilename", &matrixFilename, "If nonempty, dump the "
"generated matrix to that file in MatrixMarket format.");
// If rowMapFilename is nonempty, dump the matrix's row Map to
// that file in MatrixMarket format.
std::string rowMapFilename;
cmdp.setOption ("rowMapFilename", &rowMapFilename, "If nonempty, dump the "
"generated matrix's row Map to that file in a format that "
"Tpetra::MatrixMarket::Reader can read.");
// Option to exit after building A and b (and dumping stuff to
// files, if requested).
bool exitAfterAssembly = false;
cmdp.setOption ("exitAfterAssembly", "dontExitAfterAssembly",
&exitAfterAssembly, "If true, exit after building the "
"sparse matrix and dense right-hand side vector. If either"
" --matrixFilename or --rowMapFilename are nonempty strings"
", dump the matrix resp. row Map to their respective files "
"before exiting.");
parseCommandLineArguments (cmdp, printedHelp, argc, argv, nx, ny, nz,
xmlInputParamsFile, solverName, verbose, debug);
if (printedHelp) {
// The user specified --help at the command line to print help
// with command-line arguments. We printed help already, so quit
// with a happy return code.
return EXIT_SUCCESS;
}
setMaterialTensorOffDiagonalValue (materialTensorOffDiagonalValue);
// Both streams only print on MPI Rank 0. "out" only prints if the
// user specified --verbose.
RCP<FancyOStream> out =
getFancyOStream (rcpFromRef ((myRank == 0 && verbose) ? std::cout : blackHole));
RCP<FancyOStream> err =
getFancyOStream (rcpFromRef ((myRank == 0 && debug) ? std::cerr : blackHole));
#ifdef HAVE_MPI
*out << "PARALLEL executable" << endl;
#else
*out << "SERIAL executable" << endl;
#endif
/**********************************************************************************/
/********************************** GET XML INPUTS ********************************/
/**********************************************************************************/
ParameterList inputList;
if (xmlInputParamsFile != "") {
*out << "Reading parameters from XML file \""
<< xmlInputParamsFile << "\"..." << endl;
Teuchos::updateParametersFromXmlFile (xmlInputParamsFile,
outArg (inputList));
if (myRank == 0) {
inputList.print (*out, 2, true, true);
*out << endl;
}
}
// Get Pamgen mesh definition string, either from the input
// ParameterList or from our function that makes a cube and fills in
// the number of cells along each dimension.
std::string meshInput = inputList.get("meshInput", "");
if (meshInput == "") {
*out << "Generating mesh input string: nx = " << nx
<< ", ny = " << ny
<< ", nz = " << nz << endl;
meshInput = makeMeshInput (nx, ny, nz);
}
// Total application run time
{
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total Time", total_time);
RCP<sparse_matrix_type> A;
RCP<vector_type> B, X_exact, X;
{
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total Assembly", total_assembly);
makeMatrixAndRightHandSide (A, B, X_exact, X, comm, node, meshInput,
out, err, verbose, debug);
}
// Optionally dump the matrix and/or its row Map to files.
{
typedef Tpetra::MatrixMarket::Writer<sparse_matrix_type> writer_type;
if (matrixFilename != "") {
writer_type::writeSparseFile (matrixFilename, A);
}
if (rowMapFilename != "") {
writer_type::writeMapFile (rowMapFilename, * (A->getRowMap ()));
}
}
if (exitAfterAssembly) {
// Users might still be interested in assembly time.
Teuchos::TimeMonitor::report (comm.ptr (), std::cout);
return EXIT_SUCCESS;
}
const std::vector<MT> norms = exactResidualNorm (A, B, X_exact);
// X_exact is the exact solution of the PDE, projected onto the
// discrete mesh. It may not necessarily equal the exact solution
// of the linear system.
*out << "||B - A*X_exact||_2 = " << norms[0] << endl
<< "||B||_2 = " << norms[1] << endl
<< "||A||_F = " << norms[2] << endl;
// Setup preconditioner
std::string prec_type = inputList.get ("Preconditioner", "None");
RCP<operator_type> M;
{
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total Preconditioner Setup", total_prec);
if (prec_type == "MueLu") {
#ifdef HAVE_TRILINOSCOUPLINGS_MUELU
if (inputList.isSublist("MueLu")) {
ParameterList mueluParams = inputList.sublist("MueLu");
M = MueLu::CreateTpetraPreconditioner<ST,LO,GO,Node>(A,mueluParams);
} else {
M = MueLu::CreateTpetraPreconditioner<ST,LO,GO,Node>(A);
}
#else // NOT HAVE_TRILINOSCOUPLINGS_MUELU
TEUCHOS_TEST_FOR_EXCEPTION(
prec_type == "MueLu", std::runtime_error, "Tpetra scaling example: "
"In order to precondition with MueLu, you must have built Trilinos "
"with the MueLu package enabled.");
#endif // HAVE_TRILINOSCOUPLINGS_MUELU
}
} // setup preconditioner
// Get the convergence tolerance for each linear solve.
// If the user provided a nonnegative value at the command
// line, it overrides any value in the input ParameterList.
MT tol = STM::squareroot (STM::eps ()); // default value
if (tolFromCmdLine < STM::zero ()) {
tol = inputList.get ("Convergence Tolerance", tol);
} else {
tol = tolFromCmdLine;
}
// Get the maximum number of iterations for each linear solve.
// If the user provided a value other than -1 at the command
// line, it overrides any value in the input ParameterList.
int maxNumIters = 200; // default value
if (maxNumItersFromCmdLine == -1) {
maxNumIters = inputList.get ("Maximum Iterations", maxNumIters);
} else {
maxNumIters = maxNumItersFromCmdLine;
}
// Get the number of "time steps." We imitate a time-dependent
// PDE by doing this many linear solves.
const int num_steps = inputList.get ("Number of Time Steps", 1);
// Do the linear solve(s).
bool converged = false;
int numItersPerformed = 0;
if (gpu) {
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total GPU Solve", total_solve);
solveWithBelosGPU (converged, numItersPerformed, tol, maxNumIters,
num_steps, ranks_per_node, gpu_ranks_per_node,
device_offset, prec_type, X, A, B, Teuchos::null, M);
}
else {
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total Solve", total_solve);
solveWithBelos (converged, numItersPerformed, solverName, tol,
maxNumIters, num_steps, X, A, B, Teuchos::null, M);
}
// Compute ||X-X_exact||_2
const MT norm_x = X_exact->norm2 ();
X_exact->update (-1.0, *X, 1.0);
const MT norm_error = X_exact->norm2 ();
*out << endl
<< "||X - X_exact||_2 / ||X_exact||_2 = " << norm_error / norm_x
<< endl;
} // total time block
// Summarize timings
Teuchos::TimeMonitor::report (comm.ptr (), std::cout);
} // try
TEUCHOS_STANDARD_CATCH_STATEMENTS(true, std::cerr, success);
if (success) {
return EXIT_SUCCESS;
} else {
return EXIT_FAILURE;
}
}
int
main (int argc, char *argv[])
{
using namespace TrilinosCouplings; // Yes, this means I'm lazy.
using TpetraIntrepidPoissonExample::exactResidualNorm;
using TpetraIntrepidPoissonExample::makeMatrixAndRightHandSide;
using TpetraIntrepidPoissonExample::solveWithBelos;
using TpetraIntrepidPoissonExample::solveWithBelosGPU;
using IntrepidPoissonExample::makeMeshInput;
using IntrepidPoissonExample::setCommandLineArgumentDefaults;
using IntrepidPoissonExample::setUpCommandLineArguments;
using IntrepidPoissonExample::parseCommandLineArguments;
using Tpetra::DefaultPlatform;
using Teuchos::Comm;
using Teuchos::outArg;
using Teuchos::ParameterList;
using Teuchos::parameterList;
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcpFromRef;
using Teuchos::getFancyOStream;
using Teuchos::FancyOStream;
using std::endl;
// Pull in typedefs from the example's namespace.
typedef TpetraIntrepidPoissonExample::ST ST;
typedef TpetraIntrepidPoissonExample::LO LO;
typedef TpetraIntrepidPoissonExample::GO GO;
typedef TpetraIntrepidPoissonExample::Node Node;
typedef Teuchos::ScalarTraits<ST> STS;
typedef STS::magnitudeType MT;
typedef Teuchos::ScalarTraits<MT> STM;
typedef TpetraIntrepidPoissonExample::sparse_matrix_type sparse_matrix_type;
typedef TpetraIntrepidPoissonExample::vector_type vector_type;
typedef TpetraIntrepidPoissonExample::operator_type operator_type;
bool success = true;
try {
Teuchos::oblackholestream blackHole;
Teuchos::GlobalMPISession mpiSession (&argc, &argv, &blackHole);
const int myRank = mpiSession.getRank ();
//const int numProcs = mpiSession.getNProc ();
// Get the default communicator and Kokkos Node instance
RCP<const Comm<int> > comm =
DefaultPlatform::getDefaultPlatform ().getComm ();
RCP<Node> node = DefaultPlatform::getDefaultPlatform ().getNode ();
// Did the user specify --help at the command line to print help
// with command-line arguments?
bool printedHelp = false;
// Values of command-line arguments.
int nx, ny, nz;
std::string xmlInputParamsFile;
bool verbose, debug;
// Set default values of command-line arguments.
setCommandLineArgumentDefaults (nx, ny, nz, xmlInputParamsFile,
verbose, debug);
// Parse and validate command-line arguments.
Teuchos::CommandLineProcessor cmdp (false, true);
setUpCommandLineArguments (cmdp, nx, ny, nz, xmlInputParamsFile,
verbose, debug);
bool gpu = false;
cmdp.setOption ("gpu", "no-gpu", &gpu,
"Run example using GPU node (if supported)");
int ranks_per_node = 1;
cmdp.setOption("ranks_per_node", &ranks_per_node,
"Number of MPI ranks per node");
int gpu_ranks_per_node = 1;
cmdp.setOption("gpu_ranks_per_node", &gpu_ranks_per_node,
"Number of MPI ranks per node for GPUs");
int device_offset = 0;
cmdp.setOption("device_offset", &device_offset,
"Offset for attaching MPI ranks to CUDA devices");
parseCommandLineArguments (cmdp, printedHelp, argc, argv, nx, ny, nz,
xmlInputParamsFile, verbose, debug);
if (printedHelp) {
// The user specified --help at the command line to print help
// with command-line arguments. We printed help already, so quit
// with a happy return code.
return EXIT_SUCCESS;
}
// Both streams only print on MPI Rank 0. "out" only prints if the
// user specified --verbose.
RCP<FancyOStream> out =
getFancyOStream (rcpFromRef ((myRank == 0 && verbose) ? std::cout : blackHole));
RCP<FancyOStream> err =
getFancyOStream (rcpFromRef ((myRank == 0 && debug) ? std::cerr : blackHole));
#ifdef HAVE_MPI
*out << "PARALLEL executable" << endl;
#else
*out << "SERIAL executable" << endl;
#endif
/**********************************************************************************/
/********************************** GET XML INPUTS ********************************/
/**********************************************************************************/
ParameterList inputList;
if (xmlInputParamsFile != "") {
*out << "Reading parameters from XML file \""
<< xmlInputParamsFile << "\"..." << endl;
Teuchos::updateParametersFromXmlFile (xmlInputParamsFile,
outArg (inputList));
if (myRank == 0) {
inputList.print (*out, 2, true, true);
*out << endl;
}
}
// Get Pamgen mesh definition string, either from the input
// ParameterList or from our function that makes a cube and fills in
// the number of cells along each dimension.
std::string meshInput = inputList.get("meshInput", "");
if (meshInput == "") {
*out << "Generating mesh input string: nx = " << nx
<< ", ny = " << ny
<< ", nz = " << nz << endl;
meshInput = makeMeshInput (nx, ny, nz);
}
// Total application run time
{
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total Time", total_time);
RCP<sparse_matrix_type> A;
RCP<vector_type> B, X_exact, X;
{
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total Assembly", total_assembly);
makeMatrixAndRightHandSide (A, B, X_exact, X, comm, node, meshInput,
out, err, verbose, debug);
}
const std::vector<MT> norms = exactResidualNorm (A, B, X_exact);
// X_exact is the exact solution of the PDE, projected onto the
// discrete mesh. It may not necessarily equal the exact solution
// of the linear system.
*out << "||B - A*X_exact||_2 = " << norms[0] << endl
<< "||B||_2 = " << norms[1] << endl
<< "||A||_F = " << norms[2] << endl;
// Setup preconditioner
std::string prec_type = inputList.get("Preconditioner", "None");
RCP<operator_type> M;
{
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total Preconditioner Setup", total_prec);
if (prec_type == "MueLu") {
if (inputList.isSublist("MueLu")) {
ParameterList mueluParams = inputList.sublist("MueLu");
M = MueLu::CreateTpetraPreconditioner<ST,LO,GO,Node>(A,mueluParams);
} else {
M = MueLu::CreateTpetraPreconditioner<ST,LO,GO,Node>(A);
}
}
}
bool converged = false;
int numItersPerformed = 0;
const MT tol = inputList.get("Convergence Tolerance",
STM::squareroot (STM::eps ()));
const int maxNumIters = inputList.get("Maximum Iterations", 200);
const int num_steps = inputList.get("Number of Time Steps", 1);
if (gpu) {
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total GPU Solve", total_solve);
solveWithBelosGPU(converged, numItersPerformed, tol, maxNumIters, num_steps,
ranks_per_node, gpu_ranks_per_node, device_offset,
prec_type,
X, A, B, Teuchos::null, M);
}
else {
TEUCHOS_FUNC_TIME_MONITOR_DIFF("Total Solve", total_solve);
solveWithBelos (converged, numItersPerformed, tol, maxNumIters, num_steps,
X, A, B, Teuchos::null, M);
}
// Compute ||X-X_exact||_2
MT norm_x = X_exact->norm2();
X_exact->update(-1.0, *X, 1.0);
MT norm_error = X_exact->norm2();
*out << endl
<< "||X-X_exact||_2 / ||X_exact||_2 = " << norm_error / norm_x
<< endl;
} // total time block
// Summarize timings
// RCP<ParameterList> reportParams = parameterList ("TimeMonitor::report");
// reportParams->set ("Report format", std::string ("YAML"));
// reportParams->set ("writeGlobalStats", true);
// Teuchos::TimeMonitor::report (*out, reportParams);
Teuchos::TimeMonitor::summarize(std::cout);
} //try
TEUCHOS_STANDARD_CATCH_STATEMENTS(true, std::cerr, success);
if (success)
return EXIT_SUCCESS;
return EXIT_FAILURE;
}
