void LinearOpScalarProd<Scalar>::scalarProdsImpl(
const MultiVectorBase<Scalar>& X, const MultiVectorBase<Scalar>& Y,
const ArrayView<Scalar> &scalarProds_out
) const
{
Teuchos::RCP<MultiVectorBase<Scalar> >
T = createMembers(Y.range() ,Y.domain()->dim());
Thyra::apply(*op_, NOTRANS,Y, T.ptr());
dots(X, *T, scalarProds_out);
}
void SpmdMultiVectorSerializer<Scalar>::serialize(
const MultiVectorBase<Scalar>& mv, std::ostream& out
) const
{
Teuchos::RCP<const SpmdVectorSpaceBase<Scalar> >
mpi_vec_spc
= Teuchos::rcp_dynamic_cast<const SpmdVectorSpaceBase<Scalar> >(mv.range());
std::ios::fmtflags fmt(out.flags());
out.precision(std::numeric_limits<Scalar>::digits10+4);
if( mpi_vec_spc.get() ) {
// This is a mpi-based vector space so let's just write the local
// multi-vector elements (row-by-row).
const Ordinal
localOffset = mpi_vec_spc->localOffset(),
localSubDim = mpi_vec_spc->localSubDim();
const Range1D localRng( localOffset, localOffset+localSubDim-1 );
ConstDetachedMultiVectorView<Scalar> local_mv(mv,localRng,Range1D());
out << localSubDim << " " << local_mv.numSubCols() << std::endl;
if( binaryMode() ) {
// Write column-wise for better cache performance
for( Ordinal j = 0; j < local_mv.numSubCols(); ++j )
out.write( reinterpret_cast<const char*>(&local_mv(0,j)), sizeof(Scalar)*localSubDim );
}
else {
// Write row-wise for better readability
for( Ordinal i = 0; i < localSubDim; ++i ) {
out << " " << i;
for( Ordinal j = 0; j < local_mv.numSubCols(); ++j ) {
out << " " << local_mv(i,j);
}
out << std::endl;
}
}
}
else {
// This is a serial (or locally replicated) vector space so
// just write all of the multi-vector elements here.
TEUCHOS_TEST_FOR_EXCEPTION( true, std::logic_error, "Does not handle non-SPMD spaces yet" );
}
out.flags(fmt);
}
void doExplicitMultiVectorAdjoint(
const MultiVectorBase<Scalar>& mvIn, MultiVectorBase<Scalar>* mvTransOut
)
{
typedef Teuchos::ScalarTraits<Scalar> ST;
#ifdef TEUCHOS_DEBUG
TEST_FOR_EXCEPT(0==mvTransOut);
THYRA_ASSERT_VEC_SPACES("doExplicitMultiVectorAdjoint(...)",
*mvIn.domain(), *mvTransOut->range()
);
THYRA_ASSERT_VEC_SPACES("doExplicitMultiVectorAdjoint(...)",
*mvIn.range(), *mvTransOut->domain()
);
#endif
ConstDetachedMultiVectorView<Scalar> dMvIn(mvIn);
DetachedMultiVectorView<Scalar> dMvTransOut(*mvTransOut);
const int m = dMvIn.subDim();
const int n = dMvIn.numSubCols();
for ( int j = 0; j < n; ++j ) {
for ( int i = 0; i < m; ++i ) {
dMvTransOut(j,i) = ST::conjugate(dMvIn(i,j));
}
}
}
SolveStatus<Scalar>
BelosLinearOpWithSolve<Scalar>::solveImpl(
const EOpTransp M_trans,
const MultiVectorBase<Scalar> &B,
const Ptr<MultiVectorBase<Scalar> > &X,
const Ptr<const SolveCriteria<Scalar> > solveCriteria
) const
{
TEUCHOS_FUNC_TIME_MONITOR("BelosLOWS");
using Teuchos::rcp;
using Teuchos::rcpFromRef;
using Teuchos::rcpFromPtr;
using Teuchos::FancyOStream;
using Teuchos::OSTab;
using Teuchos::describe;
typedef Teuchos::ScalarTraits<Scalar> ST;
typedef typename ST::magnitudeType ScalarMag;
Teuchos::Time totalTimer(""), timer("");
totalTimer.start(true);
assertSolveSupports(*this, M_trans, solveCriteria);
// 2010/08/22: rabartl: Bug 4915 ToDo: Move the above into the NIV function
// solve(...).
const int numRhs = B.domain()->dim();
const int numEquations = B.range()->dim();
const RCP<FancyOStream> out = this->getOStream();
const Teuchos::EVerbosityLevel verbLevel = this->getVerbLevel();
OSTab tab = this->getOSTab();
if (out.get() && static_cast<int>(verbLevel) > static_cast<int>(Teuchos::VERB_LOW)) {
*out << "\nStarting iterations with Belos:\n";
OSTab tab2(out);
*out << "Using forward operator = " << describe(*fwdOpSrc_->getOp(),verbLevel);
*out << "Using iterative solver = " << describe(*iterativeSolver_,verbLevel);
*out << "With #Eqns="<<numEquations<<", #RHSs="<<numRhs<<" ...\n";
}
//
// Set RHS and LHS
//
bool ret = lp_->setProblem( rcpFromPtr(X), rcpFromRef(B) );
TEST_FOR_EXCEPTION(
ret == false, CatastrophicSolveFailure
,"Error, the Belos::LinearProblem could not be set for the current solve!"
);
//
// Set the solution criteria
//
const RCP<Teuchos::ParameterList> tmpPL = Teuchos::parameterList();
SolveMeasureType solveMeasureType;
RCP<GeneralSolveCriteriaBelosStatusTest<Scalar> > generalSolveCriteriaBelosStatusTest;
if (nonnull(solveCriteria)) {
solveMeasureType = solveCriteria->solveMeasureType;
const ScalarMag requestedTol = solveCriteria->requestedTol;
if (solveMeasureType.useDefault()) {
tmpPL->set("Convergence Tolerance", defaultTol_);
}
else if (solveMeasureType(SOLVE_MEASURE_NORM_RESIDUAL, SOLVE_MEASURE_NORM_RHS)) {
if (requestedTol != SolveCriteria<Scalar>::unspecifiedTolerance()) {
tmpPL->set("Convergence Tolerance", requestedTol);
}
else {
tmpPL->set("Convergence Tolerance", defaultTol_);
}
tmpPL->set("Explicit Residual Scaling", "Norm of RHS");
}
else if (solveMeasureType(SOLVE_MEASURE_NORM_RESIDUAL, SOLVE_MEASURE_NORM_INIT_RESIDUAL)) {
if (requestedTol != SolveCriteria<Scalar>::unspecifiedTolerance()) {
tmpPL->set("Convergence Tolerance", requestedTol);
}
else {
tmpPL->set("Convergence Tolerance", defaultTol_);
}
tmpPL->set("Explicit Residual Scaling", "Norm of Initial Residual");
}
else {
// Set the most generic (and inefficient) solve criteria
generalSolveCriteriaBelosStatusTest = createGeneralSolveCriteriaBelosStatusTest(
*solveCriteria, convergenceTestFrequency_);
// Set the verbosity level (one level down)
generalSolveCriteriaBelosStatusTest->setOStream(out);
generalSolveCriteriaBelosStatusTest->setVerbLevel(incrVerbLevel(verbLevel, -1));
// Set the default convergence tolerance to always converged to allow
// the above status test to control things.
tmpPL->set("Convergence Tolerance", 1.0);
}
}
else {
// No solveCriteria was even passed in!
tmpPL->set("Convergence Tolerance", defaultTol_);
}
//
// Reset the blocksize if we adding more vectors than half the number of equations,
// orthogonalization will fail on the first iteration!
//
RCP<const Teuchos::ParameterList> solverParams = iterativeSolver_->getCurrentParameters();
const int currBlockSize = Teuchos::getParameter<int>(*solverParams, "Block Size");
bool isNumBlocks = false;
int currNumBlocks = 0;
if (Teuchos::isParameterType<int>(*solverParams, "Num Blocks")) {
currNumBlocks = Teuchos::getParameter<int>(*solverParams, "Num Blocks");
isNumBlocks = true;
}
const int newBlockSize = TEUCHOS_MIN(currBlockSize,numEquations/2);
if (nonnull(out)
&& static_cast<int>(verbLevel) > static_cast<int>(Teuchos::VERB_NONE)
&& newBlockSize != currBlockSize)
{
*out << "\nAdjusted block size = " << newBlockSize << "\n";
}
//
tmpPL->set("Block Size",newBlockSize);
//
// Set the number of Krylov blocks if we are using a GMRES solver, or a solver
// that recognizes "Num Blocks". Otherwise the solver will throw an error!
//
if (isNumBlocks) {
const int Krylov_length = (currNumBlocks*currBlockSize)/newBlockSize;
tmpPL->set("Num Blocks",Krylov_length);
if (newBlockSize != currBlockSize) {
if (out.get() && static_cast<int>(verbLevel) > static_cast<int>(Teuchos::VERB_NONE))
*out
<< "\nAdjusted max number of Krylov basis blocks = " << Krylov_length << "\n";
}
}
//
// Solve the linear system
//
Belos::ReturnType belosSolveStatus;
{
RCP<std::ostream>
outUsed =
( static_cast<int>(verbLevel) > static_cast<int>(Teuchos::VERB_NONE)
? out
: rcp(new FancyOStream(rcp(new Teuchos::oblackholestream())))
);
Teuchos::OSTab tab(outUsed,1,"BELOS");
tmpPL->set("Output Stream", outUsed);
iterativeSolver_->setParameters(tmpPL);
if (nonnull(generalSolveCriteriaBelosStatusTest)) {
iterativeSolver_->setUserConvStatusTest(generalSolveCriteriaBelosStatusTest);
}
belosSolveStatus = iterativeSolver_->solve();
}
//
// Report the solve status
//
totalTimer.stop();
SolveStatus<Scalar> solveStatus;
switch (belosSolveStatus) {
case Belos::Unconverged: {
solveStatus.solveStatus = SOLVE_STATUS_UNCONVERGED;
break;
}
case Belos::Converged: {
solveStatus.solveStatus = SOLVE_STATUS_CONVERGED;
if (nonnull(generalSolveCriteriaBelosStatusTest)) {
const ArrayView<const ScalarMag> achievedTol =
generalSolveCriteriaBelosStatusTest->achievedTol();
solveStatus.achievedTol = ST::zero();
for (Ordinal i = 0; i < achievedTol.size(); ++i) {
solveStatus.achievedTol = std::max(solveStatus.achievedTol, achievedTol[i]);
}
}
else {
solveStatus.achievedTol = tmpPL->get("Convergence Tolerance", defaultTol_);
}
break;
}
TEUCHOS_SWITCH_DEFAULT_DEBUG_ASSERT();
}
std::ostringstream ossmessage;
ossmessage
<< "The Belos solver of type \""<<iterativeSolver_->description()
<<"\" returned a solve status of \""<< toString(solveStatus.solveStatus) << "\""
<< " in " << iterativeSolver_->getNumIters() << " iterations"
<< " with total CPU time of " << totalTimer.totalElapsedTime() << " sec" ;
if (out.get() && static_cast<int>(verbLevel) >=static_cast<int>(Teuchos::VERB_LOW))
*out << "\n" << ossmessage.str() << "\n";
solveStatus.message = ossmessage.str();
if (out.get() && static_cast<int>(verbLevel) >= static_cast<int>(Teuchos::VERB_LOW))
*out << "\nTotal solve time in Belos = "<<totalTimer.totalElapsedTime()<<" sec\n";
return solveStatus;
}
SolveStatus<Scalar>
BelosLinearOpWithSolve<Scalar>::solveImpl(
const EOpTransp M_trans,
const MultiVectorBase<Scalar> &B,
const Ptr<MultiVectorBase<Scalar> > &X,
const Ptr<const SolveCriteria<Scalar> > solveCriteria
) const
{
THYRA_FUNC_TIME_MONITOR("Stratimikos: BelosLOWS");
using Teuchos::rcp;
using Teuchos::rcpFromRef;
using Teuchos::rcpFromPtr;
using Teuchos::FancyOStream;
using Teuchos::OSTab;
using Teuchos::ParameterList;
using Teuchos::parameterList;
using Teuchos::describe;
typedef Teuchos::ScalarTraits<Scalar> ST;
typedef typename ST::magnitudeType ScalarMag;
Teuchos::Time totalTimer(""), timer("");
totalTimer.start(true);
assertSolveSupports(*this, M_trans, solveCriteria);
// 2010/08/22: rabartl: Bug 4915 ToDo: Move the above into the NIV function
// solve(...).
const RCP<FancyOStream> out = this->getOStream();
const Teuchos::EVerbosityLevel verbLevel = this->getVerbLevel();
OSTab tab = this->getOSTab();
if (out.get() && static_cast<int>(verbLevel) > static_cast<int>(Teuchos::VERB_LOW)) {
*out << "\nStarting iterations with Belos:\n";
OSTab tab2(out);
*out << "Using forward operator = " << describe(*fwdOpSrc_->getOp(),verbLevel);
*out << "Using iterative solver = " << describe(*iterativeSolver_,verbLevel);
*out << "With #Eqns="<<B.range()->dim()<<", #RHSs="<<B.domain()->dim()<<" ...\n";
}
//
// Set RHS and LHS
//
bool ret = lp_->setProblem( rcpFromPtr(X), rcpFromRef(B) );
TEUCHOS_TEST_FOR_EXCEPTION(
ret == false, CatastrophicSolveFailure
,"Error, the Belos::LinearProblem could not be set for the current solve!"
);
//
// Set the solution criteria
//
// Parameter list for the current solve.
const RCP<ParameterList> tmpPL = Teuchos::parameterList();
// The solver's valid parameter list.
RCP<const ParameterList> validPL = iterativeSolver_->getValidParameters();
SolveMeasureType solveMeasureType;
RCP<GeneralSolveCriteriaBelosStatusTest<Scalar> > generalSolveCriteriaBelosStatusTest;
if (nonnull(solveCriteria)) {
solveMeasureType = solveCriteria->solveMeasureType;
const ScalarMag requestedTol = solveCriteria->requestedTol;
if (solveMeasureType.useDefault()) {
tmpPL->set("Convergence Tolerance", defaultTol_);
}
else if (solveMeasureType(SOLVE_MEASURE_NORM_RESIDUAL, SOLVE_MEASURE_NORM_RHS)) {
if (requestedTol != SolveCriteria<Scalar>::unspecifiedTolerance()) {
tmpPL->set("Convergence Tolerance", requestedTol);
}
else {
tmpPL->set("Convergence Tolerance", defaultTol_);
}
setResidualScalingType (tmpPL, validPL, "Norm of RHS");
}
else if (solveMeasureType(SOLVE_MEASURE_NORM_RESIDUAL, SOLVE_MEASURE_NORM_INIT_RESIDUAL)) {
if (requestedTol != SolveCriteria<Scalar>::unspecifiedTolerance()) {
tmpPL->set("Convergence Tolerance", requestedTol);
}
else {
tmpPL->set("Convergence Tolerance", defaultTol_);
}
setResidualScalingType (tmpPL, validPL, "Norm of Initial Residual");
}
else {
// Set the most generic (and inefficient) solve criteria
generalSolveCriteriaBelosStatusTest = createGeneralSolveCriteriaBelosStatusTest(
*solveCriteria, convergenceTestFrequency_);
// Set the verbosity level (one level down)
generalSolveCriteriaBelosStatusTest->setOStream(out);
generalSolveCriteriaBelosStatusTest->setVerbLevel(incrVerbLevel(verbLevel, -1));
// Set the default convergence tolerance to always converged to allow
// the above status test to control things.
tmpPL->set("Convergence Tolerance", 1.0);
}
// maximum iterations
if (nonnull(solveCriteria->extraParameters)) {
if (Teuchos::isParameterType<int>(*solveCriteria->extraParameters,"Maximum Iterations")) {
tmpPL->set("Maximum Iterations", Teuchos::get<int>(*solveCriteria->extraParameters,"Maximum Iterations"));
}
}
}
else {
// No solveCriteria was even passed in!
tmpPL->set("Convergence Tolerance", defaultTol_);
}
//
// Solve the linear system
//
Belos::ReturnType belosSolveStatus;
{
RCP<std::ostream>
outUsed =
( static_cast<int>(verbLevel) > static_cast<int>(Teuchos::VERB_LOW)
? out
: rcp(new FancyOStream(rcp(new Teuchos::oblackholestream())))
);
Teuchos::OSTab tab1(outUsed,1,"BELOS");
tmpPL->set("Output Stream", outUsed);
iterativeSolver_->setParameters(tmpPL);
if (nonnull(generalSolveCriteriaBelosStatusTest)) {
iterativeSolver_->setUserConvStatusTest(generalSolveCriteriaBelosStatusTest);
}
belosSolveStatus = iterativeSolver_->solve();
}
//
// Report the solve status
//
totalTimer.stop();
SolveStatus<Scalar> solveStatus;
switch (belosSolveStatus) {
case Belos::Unconverged: {
solveStatus.solveStatus = SOLVE_STATUS_UNCONVERGED;
// Set achievedTol even if the solver did not converge. This is
// helpful for things like nonlinear solvers, which might be
// able to use a partially converged result, and which would
// like to know the achieved convergence tolerance for use in
// computing bounds. It's also helpful for estimating whether a
// small increase in the maximum iteration count might be
// helpful next time.
try {
// Some solvers might not have implemented achievedTol().
// The default implementation throws std::runtime_error.
solveStatus.achievedTol = iterativeSolver_->achievedTol();
} catch (std::runtime_error&) {
// Do nothing; use the default value of achievedTol.
}
break;
}
case Belos::Converged: {
solveStatus.solveStatus = SOLVE_STATUS_CONVERGED;
if (nonnull(generalSolveCriteriaBelosStatusTest)) {
// The user set a custom status test. This means that we
// should ask the custom status test itself, rather than the
// Belos solver, what the final achieved convergence tolerance
// was.
const ArrayView<const ScalarMag> achievedTol =
generalSolveCriteriaBelosStatusTest->achievedTol();
solveStatus.achievedTol = ST::zero();
for (Ordinal i = 0; i < achievedTol.size(); ++i) {
solveStatus.achievedTol = std::max(solveStatus.achievedTol, achievedTol[i]);
}
}
else {
try {
// Some solvers might not have implemented achievedTol().
// The default implementation throws std::runtime_error.
solveStatus.achievedTol = iterativeSolver_->achievedTol();
} catch (std::runtime_error&) {
// Use the default convergence tolerance. This is a correct
// upper bound, since we did actually converge.
solveStatus.achievedTol = tmpPL->get("Convergence Tolerance", defaultTol_);
}
}
break;
}
TEUCHOS_SWITCH_DEFAULT_DEBUG_ASSERT();
}
std::ostringstream ossmessage;
ossmessage
<< "The Belos solver of type \""<<iterativeSolver_->description()
<<"\" returned a solve status of \""<< toString(solveStatus.solveStatus) << "\""
<< " in " << iterativeSolver_->getNumIters() << " iterations"
<< " with total CPU time of " << totalTimer.totalElapsedTime() << " sec" ;
if (out.get() && static_cast<int>(verbLevel) > static_cast<int>(Teuchos::VERB_NONE))
*out << "\n" << ossmessage.str() << "\n";
solveStatus.message = ossmessage.str();
// Dump the getNumIters() and the achieved convergence tolerance
// into solveStatus.extraParameters, as the "Belos/Iteration Count"
// resp. "Belos/Achieved Tolerance" parameters.
if (solveStatus.extraParameters.is_null()) {
solveStatus.extraParameters = parameterList ();
}
solveStatus.extraParameters->set ("Belos/Iteration Count",
iterativeSolver_->getNumIters());\
// package independent version of the same
solveStatus.extraParameters->set ("Iteration Count",
iterativeSolver_->getNumIters());\
// NOTE (mfh 13 Dec 2011) Though the most commonly used Belos
// solvers do implement achievedTol(), some Belos solvers currently
// do not. In the latter case, if the solver did not converge, the
// reported achievedTol() value may just be the default "invalid"
// value -1, and if the solver did converge, the reported value will
// just be the convergence tolerance (a correct upper bound).
solveStatus.extraParameters->set ("Belos/Achieved Tolerance",
solveStatus.achievedTol);
// This information is in the previous line, which is printed anytime the verbosity
// is not set to Teuchos::VERB_NONE, so I'm commenting this out for now.
// if (out.get() && static_cast<int>(verbLevel) > static_cast<int>(Teuchos::VERB_NONE))
// *out << "\nTotal solve time in Belos = "<<totalTimer.totalElapsedTime()<<" sec\n";
return solveStatus;
}
bool SpmdMultiVectorSerializer<Scalar>::isCompatible(
const MultiVectorBase<Scalar> &mv
) const
{
return 0!=dynamic_cast<const SpmdVectorSpaceBase<Scalar>*>(&*mv.range());
}
void SpmdMultiVectorBase<Scalar>::euclideanApply(
const EOpTransp M_trans,
const MultiVectorBase<Scalar> &X,
const Ptr<MultiVectorBase<Scalar> > &Y,
const Scalar alpha,
const Scalar beta
) const
{
typedef Teuchos::ScalarTraits<Scalar> ST;
using Teuchos::Workspace;
Teuchos::WorkspaceStore* wss = Teuchos::get_default_workspace_store().get();
#ifdef THYRA_SPMD_MULTI_VECTOR_BASE_PRINT_TIMES
Teuchos::Time timerTotal("dummy",true);
Teuchos::Time timer("dummy");
#endif
//
// This function performs one of two operations.
//
// The first operation (M_trans == NOTRANS) is:
//
// Y = beta * Y + alpha * M * X
//
// where Y and M have compatible (distributed?) range vector
// spaces and X is a locally replicated serial multi-vector. This
// operation does not require any global communication.
//
// The second operation (M_trans == TRANS) is:
//
// Y = beta * Y + alpha * M' * X
//
// where M and X have compatible (distributed?) range vector spaces
// and Y is a locally replicated serial multi-vector. This operation
// requires a local reduction.
//
//
// Get spaces and validate compatibility
//
// Get the SpmdVectorSpace
const SpmdVectorSpaceBase<Scalar> &spmdSpc = *this->spmdSpace();
// Get the Spmd communicator
const RCP<const Teuchos::Comm<Ordinal> >
comm = spmdSpc.getComm();
#ifdef TEUCHOS_DEBUG
const VectorSpaceBase<Scalar>
&Y_range = *Y->range(),
&X_range = *X.range();
// std::cout << "SpmdMultiVectorBase<Scalar>::apply(...): comm = " << comm << std::endl;
TEUCHOS_TEST_FOR_EXCEPTION(
( globalDim_ > localSubDim_ ) && comm.get()==NULL, std::logic_error
,"SpmdMultiVectorBase<Scalar>::apply(...MultiVectorBase<Scalar>...): Error!"
);
// ToDo: Write a good general validation function that I can call that will replace
// all of these TEUCHOS_TEST_FOR_EXCEPTION(...) uses
TEUCHOS_TEST_FOR_EXCEPTION(
real_trans(M_trans)==NOTRANS && !spmdSpc.isCompatible(Y_range), Exceptions::IncompatibleVectorSpaces
,"SpmdMultiVectorBase<Scalar>::apply(...MultiVectorBase<Scalar>...): Error!"
);
TEUCHOS_TEST_FOR_EXCEPTION(
real_trans(M_trans)==TRANS && !spmdSpc.isCompatible(X_range), Exceptions::IncompatibleVectorSpaces
,"SpmdMultiVectorBase<Scalar>::apply(...MultiVectorBase<Scalar>...): Error!"
);
#endif
//
// Get explicit (local) views of Y, M and X
//
#ifdef THYRA_SPMD_MULTI_VECTOR_BASE_PRINT_TIMES
timer.start();
#endif
DetachedMultiVectorView<Scalar>
Y_local(
*Y,
real_trans(M_trans)==NOTRANS ? Range1D(localOffset_,localOffset_+localSubDim_-1) : Range1D(),
Range1D()
);
ConstDetachedMultiVectorView<Scalar>
M_local(
*this,
Range1D(localOffset_,localOffset_+localSubDim_-1),
Range1D()
);
ConstDetachedMultiVectorView<Scalar>
X_local(
X
,real_trans(M_trans)==NOTRANS ? Range1D() : Range1D(localOffset_,localOffset_+localSubDim_-1)
,Range1D()
);
#ifdef THYRA_SPMD_MULTI_VECTOR_BASE_PRINT_TIMES
timer.stop();
std::cout << "\nSpmdMultiVectorBase<Scalar>::apply(...): Time for getting view = " << timer.totalElapsedTime() << " seconds\n";
#endif
#ifdef TEUCHOS_DEBUG
TEUCHOS_TEST_FOR_EXCEPTION(
real_trans(M_trans)==NOTRANS && ( M_local.numSubCols() != X_local.subDim() || X_local.numSubCols() != Y_local.numSubCols() )
, Exceptions::IncompatibleVectorSpaces
,"SpmdMultiVectorBase<Scalar>::apply(...MultiVectorBase<Scalar>...): Error!"
);
TEUCHOS_TEST_FOR_EXCEPTION(
real_trans(M_trans)==TRANS && ( M_local.subDim() != X_local.subDim() || X_local.numSubCols() != Y_local.numSubCols() )
, Exceptions::IncompatibleVectorSpaces
,"SpmdMultiVectorBase<Scalar>::apply(...MultiVectorBase<Scalar>...): Error!"
);
#endif
//
// If nonlocal (i.e. M_trans==TRANS) then create temporary storage
// for:
//
// Y_local_tmp = alpha * M(local) * X(local) : on nonroot processes
//
// or
//
// Y_local_tmp = beta*Y_local + alpha * M(local) * X(local) : on root process (localOffset_==0)
//
// and set
//
// localBeta = ( localOffset_ == 0 ? beta : 0.0 )
//
// Above, we choose localBeta such that we will only perform
// Y_local = beta * Y_local + ... on one process (the root
// process where localOffset_==0x). Then, when we add up Y_local
// on all of the processors and we will get the correct result.
//
// If strictly local (i.e. M_trans == NOTRANS) then set:
//
// Y_local_tmp = Y_local
// localBeta = beta
//
#ifdef THYRA_SPMD_MULTI_VECTOR_BASE_PRINT_TIMES
timer.start();
#endif
Workspace<Scalar> Y_local_tmp_store(wss, Y_local.subDim()*Y_local.numSubCols(), false);
RTOpPack::SubMultiVectorView<Scalar> Y_local_tmp;
Scalar localBeta;
if( real_trans(M_trans) == TRANS && globalDim_ > localSubDim_ ) {
// Nonlocal
Y_local_tmp.initialize(
0, Y_local.subDim(),
0, Y_local.numSubCols(),
Teuchos::arcpFromArrayView(Y_local_tmp_store()),
Y_local.subDim() // leadingDim == subDim (columns are adjacent)
);
if( localOffset_ == 0 ) {
// Root process: Must copy Y_local into Y_local_tmp
for( int j = 0; j < Y_local.numSubCols(); ++j ) {
Scalar *Y_local_j = Y_local.values() + Y_local.leadingDim()*j;
std::copy( Y_local_j, Y_local_j + Y_local.subDim(), Y_local_tmp.values() + Y_local_tmp.leadingDim()*j );
}
localBeta = beta;
}
else {
// Not the root process
localBeta = 0.0;
}
}
else {
// Local
Y_local_tmp = Y_local.smv(); // Shallow copy only!
localBeta = beta;
}
#ifdef THYRA_SPMD_MULTI_VECTOR_BASE_PRINT_TIMES
timer.stop();
std::cout << "\nSpmdMultiVectorBase<Scalar>::apply(...): Time for setting up Y_local_tmp and localBeta = " << timer.totalElapsedTime() << " seconds\n";
#endif
//
// Perform the local multiplication:
//
// Y(local) = localBeta * Y(local) + alpha * op(M(local)) * X(local)
//
// or in BLAS lingo:
//
// C = beta * C + alpha * op(A) * op(B)
//
#ifdef THYRA_SPMD_MULTI_VECTOR_BASE_PRINT_TIMES
timer.start();
#endif
Teuchos::ETransp t_transp;
if(ST::isComplex) {
switch(M_trans) {
case NOTRANS: t_transp = Teuchos::NO_TRANS; break;
case TRANS: t_transp = Teuchos::TRANS; break;
case CONJTRANS: t_transp = Teuchos::CONJ_TRANS; break;
default: TEUCHOS_TEST_FOR_EXCEPT(true);
}
}
else {
switch(real_trans(M_trans)) {
case NOTRANS: t_transp = Teuchos::NO_TRANS; break;
case TRANS: t_transp = Teuchos::TRANS; break;
default: TEUCHOS_TEST_FOR_EXCEPT(true);
}
}
if (M_local.numSubCols() > 0) {
// AGS: Added std::max on ld? below, following what is done in
// Epetra_MultiVector Multiply use of GEMM. Allows for 0 length.
blas_.GEMM(
t_transp // TRANSA
,Teuchos::NO_TRANS // TRANSB
,Y_local.subDim() // M
,Y_local.numSubCols() // N
,real_trans(M_trans)==NOTRANS ? M_local.numSubCols() : M_local.subDim() // K
,alpha // ALPHA
,const_cast<Scalar*>(M_local.values()) // A
,std::max((int) M_local.leadingDim(),1) // LDA
,const_cast<Scalar*>(X_local.values()) // B
,std::max((int) X_local.leadingDim(),1) // LDB
,localBeta // BETA
,Y_local_tmp.values().get() // C
,std::max((int) Y_local_tmp.leadingDim(),1) // LDC
);
}
else {
std::fill( Y_local_tmp.values().begin(), Y_local_tmp.values().end(),
ST::zero() );
}
#ifdef THYRA_SPMD_MULTI_VECTOR_BASE_PRINT_TIMES
timer.stop();
std::cout
<< "\nSpmdMultiVectorBase<Scalar>::apply(...): Time for GEMM = "
<< timer.totalElapsedTime() << " seconds\n";
#endif
if( comm.get() ) {
//
// Perform the global reduction of Y_local_tmp back into Y_local
//
if( real_trans(M_trans)==TRANS && globalDim_ > localSubDim_ ) {
// Contiguous buffer for final reduction
Workspace<Scalar> Y_local_final_buff(wss,Y_local.subDim()*Y_local.numSubCols(),false);
// Perform the reduction
Teuchos::reduceAll<Ordinal,Scalar>(
*comm,Teuchos::REDUCE_SUM,Y_local_final_buff.size(),Y_local_tmp.values().get(),
&Y_local_final_buff[0]
);
// Load Y_local_final_buff back into Y_local
const Scalar *Y_local_final_buff_ptr = &Y_local_final_buff[0];
for( int j = 0; j < Y_local.numSubCols(); ++j ) {
Scalar *Y_local_ptr = Y_local.values() + Y_local.leadingDim()*j;
for( int i = 0; i < Y_local.subDim(); ++i ) {
(*Y_local_ptr++) = (*Y_local_final_buff_ptr++);
}
}
}
}
else {
// When you get here the view Y_local will be committed back to Y
// in the destructor to Y_local
}
#ifdef THYRA_SPMD_MULTI_VECTOR_BASE_PRINT_TIMES
timer.stop();
std::cout
<< "\nSpmdMultiVectorBase<Scalar>::apply(...): Total time = "
<< timerTotal.totalElapsedTime() << " seconds\n";
#endif
}
