void
dft_PolyA11_Tpetra_Operator<Scalar,MatrixType>::
Check
(bool verbose) const
{
RCP<VEC > x = rcp(new VEC(getDomainMap()));
RCP<VEC > b = rcp(new VEC(getRangeMap()));
x->randomize(); // Fill x with random numbers
apply(*x, *b); // Forward operation
applyInverse(*b, *b); // Reverse operation
b->update(-STS::one(), *x, STS::one()); // Should be zero
Scalar absResid = b->norm2();
Scalar normX = x->norm2();
Scalar resid = absResid / normX;
if (verbose)
{
std::cout << "A11 self-check residual = " << resid << std::endl;
}
TEUCHOS_TEST_FOR_EXCEPTION(resid > 1.0E-12, std::runtime_error, "Bad residual.\n");
} //end Check
void InputRange::getRangeVector(WirelessModels::NodeModel nodeType, WirelessTypes::ChannelType channelType, WirelessTypes::InputRanges& result)
{
const auto& rangeMap = getRangeMap(nodeType, channelType);
for(auto& range : rangeMap)
{
result.push_back(range.second);
}
}
static void run_performance_test (const Input& in) {
typedef LO Int;
typedef tif_utest::BlockCrsMatrixMaker<Scalar, LO, GO> bcmm;
typedef tif_utest::BlockTriDiContainerTester<Scalar, LO, GO> btdct;
// Performance test. Use BlockTriDiContainer directly.
typename bcmm::StructuredBlock sb(in.ni, in.nj, in.nk, in.contiguous);
typename bcmm::StructuredBlockPart sbp = bcmm::make_StructuredBlockPart(sb, in.isplit, in.jsplit,
in.comm->getRank());
auto g = bcmm::make_crs_graph(in.comm, sb, sbp);
auto A = bcmm::make_bcrs_matrix(sb, g, in.bs);
const auto T = btdct::make_BTDC(sb, sbp, A, in.use_overlap_comm, in.nonuniform_lines, in.jacobi,
in.use_seq_method);
const auto X = bcmm::make_multivector(sb, A, in.bs, in.nrhs);
const auto B = Teuchos::rcp(new typename bcmm::Tpetra_MultiVector(A->getRangeMap(), in.nrhs));
A->apply(*X, *B);
const auto X_solve = Teuchos::rcp(new typename bcmm::Tpetra_MultiVector(A->getDomainMap(), in.nrhs));
Teuchos::TimeMonitor::zeroOutTimers();
T->initialize();
auto input = T->createDefaultApplyParameters();
input.zeroStartingSolution = true;
input.maxNumSweeps = in.iterations;
input.tolerance = in.tol;
input.checkToleranceEvery = in.check_tol_every;
int sweep = 0;
{
TEUCHOS_FUNC_TIME_MONITOR("Performance test");
for (Int repeat = 0; repeat < in.repeat; ++repeat) {
T->compute();
sweep = T->applyInverseJacobi(*B, *X_solve, input);
}
}
const auto rd = bcmm::reldif(*X, *X_solve);
if (in.verbose && in.comm->getRank() == 0)
std::cout << "Performance test rd = " << rd << "\n";
if (in.comm->getRank() == 0) {
const auto n0 = T->getNorms0();
const auto nf = T->getNormsFinal();
bool ok = true;
if (n0.size() > 0) {
std::stringstream ss;
ss << "Norm reduction:";
for (int i = 0; i < n0.size(); ++i) {
const auto r = nf[i] / n0[i];
ss << " " << r;
if (r > in.tol && sweep != in.iterations) ok = false;
}
ss << "\n";
if (in.verbose)
std::cout << ss.str();
if ( ! ok)
std::cout << "FAIL: norm-based termination\n";
}
}
Teuchos::TimeMonitor::summarize();
}
void
dft_PolyA22_Tpetra_Operator<Scalar,MatrixType>::
Check
(bool verbose) const
{
RCP<VEC > x = rcp(new VEC(getDomainMap()));
RCP<VEC > b = rcp(new VEC(getRangeMap()));
x->randomize(); // Fill x with random numbers
apply(*x, *b); // Forward operation
if (hasDensityOnCms_)
{
if (F_location_ == 1) //F in NE
{
// Inverse is not exact, so we must modify b2 first:
RCP<MV> x1 = x->offsetViewNonConst(cmsMap_, 0);
// Start x1 to view first numCmsElements elements of x
RCP<MV> b2 = b->offsetViewNonConst(densityMap_, cmsMap_->getNodeNumElements());
// Start b2 to view last numDensity elements of b
RCP<MV > DCx1 = rcp(new MV(*b2));
densityOnCmsMatrixOp_->apply(*x1, *DCx1);
b2->update(-STS::one(), *DCx1, STS::one()); // b2 = b2 - DC*x1
}
else
{
// Inverse is not exact, so we must modify b1 first:
RCP<MV> x2 = x->offsetViewNonConst(cmsMap_, densityMap_->getNodeNumElements());
//Start x2 to view last numCms elements of x
RCP<MV> b1 = b->offsetViewNonConst(densityMap_, 0);
// Start b1 to view first numDensity elements of b
RCP<MV > DCx2 = rcp(new MV(*b1));
densityOnCmsMatrixOp_->apply(*x2, *DCx2);
b1->update(-STS::one(), *DCx2, STS::one()); // b1 = b1 - DC*x2
}
}
// Reverse operation
applyInverse(*b, *b);
b->update(-STS::one(), *x, STS::one()); // Should be zero
Scalar resid = b->norm2();
if (verbose)
{
std::cout << "A22 self-check residual = " << resid << std::endl;
}
TEUCHOS_TEST_FOR_EXCEPTION(resid > 1.0E-12, std::runtime_error, "Bad residual.\n");
} //end Check
WirelessTypes::InputRange InputRange::eepromValToInputRange(uint16 eepromValue, WirelessModels::NodeModel nodeType, WirelessTypes::ChannelType channelType)
{
const InputRangeMap& ranges = getRangeMap(nodeType, channelType);
try
{
return ranges.at(eepromValue);
}
catch(std::out_of_range&)
{
return WirelessTypes::range_invalid;
}
}
uint16 InputRange::inputRangeToEepromVal(WirelessTypes::InputRange range, WirelessModels::NodeModel nodeType, WirelessTypes::ChannelType channelType)
{
const InputRangeMap& ranges = getRangeMap(nodeType, channelType);
for(auto& r : ranges)
{
if(r.second == range)
{
return r.first;
}
}
throw Error_NotSupported("Invalid Input Range Value");
}
//TODO : Get ride of "Tpetra"::OptimizeOption
void fillComplete(const RCP<ParameterList> ¶ms=null)
{
for (size_t r=0; r<Rows(); ++r)
{
for (size_t c=0; c<Cols(); ++c)
{
if(getMatrix(r,c)->isFillComplete() == false)
getMatrix(r,c)->fillComplete(getDomainMap(c),getRangeMap(r),params);
}
}
// get full row map
//fullrowmap_ = Xpetra::MapFactory<LocalOrdinal,GlobalOrdinal,Node>::Build(rangemaps_->getFullMap()->lib(), rangemaps_->getFullMap()->getNodeNumElements(), rangemaps_->getFullMap()->getNodeElementList(), rangemaps_->getFullMap()->getIndexBase(), rangemaps_->getFullMap()->getComm());//rangemaps_->FullMap(); //->Clone();
fullrowmap_ = Xpetra::MapFactory<LocalOrdinal,GlobalOrdinal,Node>::Build(rangemaps_->getFullMap()->lib(), rangemaps_->getFullMap()->getGlobalNumElements(), rangemaps_->getFullMap()->getNodeElementList(), rangemaps_->getFullMap()->getIndexBase(), rangemaps_->getFullMap()->getComm());//rangemaps_->FullMap(); //->Clone();
// TODO: check me, clean up, use only ArrayView instead of std::vector
// build full col map
fullcolmap_ = Teuchos::null; // delete old full column map
if (fullcolmap_ == Teuchos::null)
{
std::vector<GlobalOrdinal> colmapentries;
for (size_t c=0; c<Cols(); ++c)
{
std::set<GlobalOrdinal> colset;
for (size_t r=0; r<Rows(); ++r)
{
if(getMatrix(r,c) != Teuchos::null) {
Teuchos::RCP<const Map> colmap = getMatrix(r,c)->getColMap();
copy(colmap->getNodeElementList().getRawPtr(),
colmap->getNodeElementList().getRawPtr()+colmap->getNodeNumElements(),
inserter(colset,colset.begin()));
}
}
colmapentries.reserve(colmapentries.size()+colset.size());
copy(colset.begin(), colset.end(), back_inserter(colmapentries));
sort(colmapentries.begin(), colmapentries.end());
typename std::vector<GlobalOrdinal>::iterator gendLocation;
gendLocation = std::unique(colmapentries.begin(), colmapentries.end());
colmapentries.erase(gendLocation,colmapentries.end());
}
// sum up number of local elements
size_t numGlobalElements = 0;
sumAll(rangemaps_->getFullMap()->getComm(), colmapentries.size(), numGlobalElements)
const Teuchos::ArrayView<const GlobalOrdinal> aView = Teuchos::ArrayView<const GlobalOrdinal>(colmapentries);
fullcolmap_ = Xpetra::MapFactory<LocalOrdinal,GlobalOrdinal,Node>::Build(rangemaps_->getFullMap()->lib(), numGlobalElements, aView, 0,rangemaps_->getFullMap()->getComm());
//fullcolmap_ = Xpetra::MapFactory<LocalOrdinal,GlobalOrdinal,Node>::Build(rangemaps_->getFullMap()->lib(), domainmaps_->getFullMap()->getGlobalNumElements(), aView, 0,domainmaps_->getFullMap()->getComm());
}
void
dft_HardSphereA22_Tpetra_Operator<Scalar,MatrixType>::
formA22Matrix
()
{
if (A22Matrix_ == Teuchos::null) {
A22Matrix_ = rcp(new MAT(getRangeMap(), 1));
A22Matrix_->setObjectLabel("HardSphereA22::A22Matrix");
}
LocalOrdinal numRows = getRangeMap()->getNodeNumElements();
Teuchos::ArrayRCP<const Scalar> vectorValues = densityOnDensityMatrix_->get1dView();
for (LocalOrdinal i=OTLO::zero(); i<numRows; i++) {
GlobalOrdinal row = A22Matrix_->getRowMap()->getGlobalElement(i);
Scalar value = vectorValues[i];
GlobalOrdinal col = row;
Array<GlobalOrdinal> cols(1);
Array<MatScalar> vals(1);
cols[0] = col;
vals[0] = Teuchos::as<MatScalar>(value);
A22Matrix_->insertGlobalValues(row, cols, vals);
}
A22Matrix_->fillComplete();
}
void
dft_HardSphereA22_Tpetra_Operator<Scalar,MatrixType>::
apply
(const MV& X, MV& Y, Teuchos::ETransp mode, Scalar alpha, Scalar beta) const
{
TEUCHOS_TEST_FOR_EXCEPT(Y.getNumVectors()!=X.getNumVectors());
#ifdef KDEBUG
TEUCHOS_TEST_FOR_EXCEPT(!X.getMap()->isSameAs(*getDomainMap()));
TEUCHOS_TEST_FOR_EXCEPT(!Y.getMap()->isSameAs(*getRangeMap()));
#endif
Y.elementWiseMultiply(STS::one(), *densityOnDensityMatrix_, X, STS::zero());
}
void
dft_HardSphereA22_Tpetra_Operator<Scalar,MatrixType>::
applyInverse
(const MV& X, MV& Y) const
{
// Our algorithm is:
// Y = D \ X
TEUCHOS_TEST_FOR_EXCEPT(Y.getNumVectors()!=X.getNumVectors());
#ifdef KDEBUG
TEUCHOS_TEST_FOR_EXCEPT(!X.getMap()->isSameAs(*getDomainMap()));
TEUCHOS_TEST_FOR_EXCEPT(!Y.getMap()->isSameAs(*getRangeMap()));
#endif
Y.elementWiseMultiply(STS::one(), *densityOnDensityInverse_, X, STS::zero());
}
void
dft_PolyA11_Tpetra_Operator<Scalar,MatrixType>::
applyInverse
(const MV& X, MV& Y) const
{
TEUCHOS_TEST_FOR_EXCEPT(Y.getNumVectors()!=X.getNumVectors());
#ifdef KDEBUG
TEUCHOS_TEST_FOR_EXCEPT(!X.getMap()->isSameAs(*getDomainMap()));
TEUCHOS_TEST_FOR_EXCEPT(!Y.getMap()->isSameAs(*getRangeMap()));
printf("\n\n\n\ndft_PolyA11_Tpetra_Operator::applyInverse()\n\n\n\n");
#endif
Scalar ONE = STS::one();
Scalar ZERO = STS::zero();
size_t NumVectors = Y.getNumVectors();
size_t numMyElements = ownedMap_->getNodeNumElements();
RCP<MV > Ytmp = rcp(new MV(ownedMap_,NumVectors));
Y=X; // We can safely do this
RCP<MV > curY = Y.offsetViewNonConst(ownedMap_, 0);
RCP<VEC> diagVec = invDiagonal_->offsetViewNonConst(ownedMap_, 0)->getVectorNonConst(0);
curY->elementWiseMultiply(ONE, *diagVec, *curY, ZERO); // Scale Y by the first block diagonal
// Loop over block 1 through numBlocks (indexing 0 to numBlocks-1)
for (LocalOrdinal i=OTLO::zero(); i< numBlocks_-1; i++)
{
// Update views of Y and diagonal blocks
curY = Y.offsetViewNonConst(ownedMap_, (i+1)*numMyElements);
diagVec = invDiagonal_->offsetViewNonConst(ownedMap_, (i+1)*numMyElements)->getVectorNonConst(0);
matrixOperator_[i]->apply(Y, *Ytmp); // Multiply block lower triangular block
curY->update(-ONE, *Ytmp, ONE); // curY = curX - Ytmp (Note that curX is in curY from initial copy Y = X)
curY->elementWiseMultiply(ONE, *diagVec, *curY, ZERO); // Scale Y by the first block diagonal
}
} //end applyInverse
// Returns explicit matrix representation of operator if available.
Teuchos::RCP<Tpetra_CrsMatrix>
LCM::
Schwarz_BoundaryJacobian::
getExplicitOperator() const
{
auto const
max_num_cols = getDomainMap()->getNodeNumElements();
Teuchos::RCP<Tpetra_CrsMatrix>
K = Teuchos::rcp(
new Tpetra_CrsMatrix(getRangeMap(), getDomainMap(), max_num_cols));
auto const
zero = Teuchos::ScalarTraits<ST>::zero();
K->setAllToScalar(zero);
K->fillComplete();
return K;
}
void
dft_PolyA11_Tpetra_Operator<Scalar,MatrixType>::
apply
(const MV& X, MV& Y, Teuchos::ETransp mode, Scalar alpha, Scalar beta) const
{
TEUCHOS_TEST_FOR_EXCEPT(Y.getNumVectors()!=X.getNumVectors());
#ifdef KDEBUG
TEUCHOS_TEST_FOR_EXCEPT(!X.getMap()->isSameAs(*getDomainMap()));
TEUCHOS_TEST_FOR_EXCEPT(!Y.getMap()->isSameAs(*getRangeMap()));
#endif
size_t numMyElements = ownedMap_->getNodeNumElements();
RCP<MV > curY = Y.offsetViewNonConst(ownedMap_, 0);
for (LocalOrdinal i=OTLO::zero(); i< numBlocks_-1; i++) {
curY = Y.offsetViewNonConst(ownedMap_, (i+1)*numMyElements);
matrixOperator_[i]->apply(X, *curY); // This gives a result that is off-diagonal-matrix*X
}
Y.elementWiseMultiply(STS::one(),*diagonal_, X, STS::one()); // Add diagonal contribution
} //end Apply
void
dft_PolyA22_Tpetra_Operator<Scalar,MatrixType>::
apply
(const MV& X, MV& Y, Teuchos::ETransp mode, Scalar alpha, Scalar beta) const
{
TEUCHOS_TEST_FOR_EXCEPT(Y.getNumVectors()!=X.getNumVectors());
#ifdef KDEBUG
TEUCHOS_TEST_FOR_EXCEPT(!X.getMap()->isSameAs(*getDomainMap()));
TEUCHOS_TEST_FOR_EXCEPT(!Y.getMap()->isSameAs(*getRangeMap()));
#endif
Scalar ONE = STS::one();
Scalar ZERO = STS::zero();
if (F_location_ == 1)
{
//F in NE
size_t numCmsElements = cmsMap_->getNodeNumElements();
// Y1 is a view of the first numCms elements of Y
RCP<MV> Y1 = Y.offsetViewNonConst(cmsMap_, 0);
// Y2 is a view of the last numDensity elements of Y
RCP<MV> Y2 = Y.offsetViewNonConst(densityMap_, numCmsElements);
// X1 is a view of the first numCms elements of X
RCP<const MV> X1 = X.offsetView(cmsMap_, 0);
// X2 is a view of the last numDensity elements of X
RCP<const MV> X2 = X.offsetView(densityMap_, numCmsElements);
// First block row
cmsOnDensityMatrixOp_->apply(*X2, *Y1);
cmsOnCmsMatrixOp_->apply(*X1, *tmpCmsVec_);
Y1->update(ONE, *tmpCmsVec_, ONE);
// Second block row
if (hasDensityOnCms_) {
densityOnCmsMatrixOp_->apply(*X1, *Y2);
Y2->elementWiseMultiply(ONE, *densityOnDensityMatrix_, *X2, ONE);
} else {
Y2->elementWiseMultiply(ONE, *densityOnDensityMatrix_, *X2, ZERO);
}
}
else
{
//F in SW
size_t numDensityElements = densityMap_->getNodeNumElements();
// Y1 is a view of the first numDensity elements of Y
RCP<MV> Y1 = Y.offsetViewNonConst(densityMap_, 0);
// Y2 is a view of the last numCms elements of Y
RCP<MV> Y2 = Y.offsetViewNonConst(cmsMap_, numDensityElements);
// X1 is a view of the first numDensity elements of X
RCP<const MV> X1 = X.offsetView(densityMap_, 0);
// X2 is a view of the last numCms elements of X
RCP<const MV> X2 = X.offsetView(cmsMap_, numDensityElements);
// First block row
if (hasDensityOnCms_) {
densityOnCmsMatrixOp_->apply(*X2, *Y1);
Y1->elementWiseMultiply(ONE, *densityOnDensityMatrix_, *X1, ONE);
} else {
Y1->elementWiseMultiply(ONE, *densityOnDensityMatrix_, *X1, ZERO);
}
// Second block row
cmsOnDensityMatrixOp_->apply(*X1, *Y2);
cmsOnCmsMatrixOp_->apply(*X2, *tmpCmsVec_);
Y2->update(ONE, *tmpCmsVec_, ONE);
}
} //end Apply
void
dft_PolyA22_Tpetra_Operator<Scalar,MatrixType>::
applyInverse
(const MV& X, MV& Y) const
{
// The true A22 block is of the form:
// | Dcc F |
// | Ddc Ddd |
// where
// Dcc is Cms on Cms (diagonal),
// F is Cms on Density (fairly dense)
// Ddc is Density on Cms (diagonal with small coefficient values),
// Ddd is Density on Density (diagonal).
//
// We will approximate A22 with:
// | Dcc F |
// | 0 Ddd |
// replacing Ddc with a zero matrix for the applyInverse method only.
// Our algorithm is then:
// Y2 = Ddd \ X2
// Y1 = Dcc \ (X1 - F*Y2)
// Or, if F is in the SW quadrant:
// The true A22 block is of the form:
// | Ddd Ddc |
// | F Dcc |
// where
// Ddd is Density on Density (diagonal),
// Ddc is Density on Cms (diagonal with small coefficient values),
// F is Cms on Density (fairly dense),
// Dcc is Cms on Cms (diagonal).
//
// We will approximate A22 with:
// | Ddd 0 |
// | F Dcc |
// replacing Ddc with a zero matrix for the applyInverse method only.
// Our algorithm is then:
// Y1 = Ddd \ X1
// Y2 = Dcc \ (X2 - F*Y1)
TEUCHOS_TEST_FOR_EXCEPT(Y.getNumVectors()!=X.getNumVectors());
#ifdef KDEBUG
TEUCHOS_TEST_FOR_EXCEPT(!X.getMap()->isSameAs(*getDomainMap()));
TEUCHOS_TEST_FOR_EXCEPT(!Y.getMap()->isSameAs(*getRangeMap()));
printf("\n\n\n\ndft_PolyA22_Tpetra_Operator::applyInverse()\n\n\n\n");
#endif
Scalar ONE = STS::one();
Scalar ZERO = STS::zero();
if (F_location_ == 1)
{
//F in NE
size_t numCmsElements = cmsMap_->getNodeNumElements();
// Y1 is a view of the first numCms elements of Y
RCP<MV > Y1 = Y.offsetViewNonConst(cmsMap_, 0);
// Y2 is a view of the last numDensity elements of Y
RCP<MV > Y2 = Y.offsetViewNonConst(densityMap_, numCmsElements);
// X1 is a view of the first numCms elements of X
RCP<const MV > X1 = X.offsetView(cmsMap_, 0);
// X2 is a view of the last numDensity elements of X
RCP<const MV > X2 = X.offsetView(densityMap_, numCmsElements);
// Second block row: Y2 = DD\X2
Y2->elementWiseMultiply(ONE, *densityOnDensityInverse_, *X2, ZERO);
// First block row: Y1 = CC \ (X1 - CD*Y2)
cmsOnDensityMatrixOp_->apply(*Y2, *tmpCmsVec_);
tmpCmsVec_->update(ONE, *X1, -ONE);
cmsOnCmsInverseOp_->apply(*tmpCmsVec_, *Y1);
}
else
{
//F in SW
size_t numDensityElements = densityMap_->getNodeNumElements();
// Y1 is a view of the first numDensity elements of Y
RCP<MV > Y1 = Y.offsetViewNonConst(densityMap_, 0);
// Y2 is a view of the last numCms elements of Y
RCP<MV > Y2 = Y.offsetViewNonConst(cmsMap_, numDensityElements);
// X1 is a view of the first numDensity elements of X
RCP<const MV > X1 = X.offsetView(densityMap_, 0);
// X2 is a view of the last numCms elements of X
RCP<const MV > X2 = X.offsetView(cmsMap_, numDensityElements);
// First block row: Y1 = DD\X1
Y1->elementWiseMultiply(ONE, *densityOnDensityInverse_, *X1, ZERO);
// Second block row: Y2 = CC \ (X2 - CD*Y1)
cmsOnDensityMatrixOp_->apply(*Y1, *tmpCmsVec_);
tmpCmsVec_->update(ONE, *X2, -ONE);
cmsOnCmsInverseOp_->apply(*tmpCmsVec_, *Y2);
}
} //end applyInverse
