clsparseStatus
cg(cldenseVectorPrivate *pX,
const clsparseCsrMatrixPrivate* pA,
const cldenseVectorPrivate *pB,
PTYPE& M,
clSParseSolverControl solverControl,
clsparseControl control)
{
assert( pA->num_cols == pB->num_values );
assert( pA->num_rows == pX->num_values );
if( ( pA->num_cols != pB->num_values ) || ( pA->num_rows != pX->num_values ) )
{
return clsparseInvalidSystemSize;
}
//opaque input parameters with clsparse::array type;
clsparse::vector<T> x(control, pX->values, pX->num_values);
clsparse::vector<T> b(control, pB->values, pB->num_values);
cl_int status;
T scalarOne = 1;
T scalarZero = 0;
//clsparse::vector<T> norm_b(control, 1, 0, CL_MEM_WRITE_ONLY, true);
clsparse::scalar<T> norm_b(control, 0, CL_MEM_WRITE_ONLY, false);
//norm of rhs of equation
status = Norm1<T>(norm_b, b, control);
CLSPARSE_V(status, "Norm B Failed");
//norm_b is calculated once
T h_norm_b = norm_b[0];
#ifndef NDEBUG
std::cout << "norm_b " << h_norm_b << std::endl;
#endif
if (h_norm_b == 0) //special case b is zero so solution is x = 0
{
solverControl->nIters = 0;
solverControl->absoluteTolerance = 0.0;
solverControl->relativeTolerance = 0.0;
//we can either fill the x with zeros or cpy b to x;
x = b;
return clsparseSuccess;
}
//continuing "normal" execution of cg algorithm
const auto N = pA->num_cols;
//helper containers, all need to be zeroed
clsparse::vector<T> y(control, N, 0, CL_MEM_READ_WRITE, true);
clsparse::vector<T> z(control, N, 0, CL_MEM_READ_WRITE, true);
clsparse::vector<T> r(control, N, 0, CL_MEM_READ_WRITE, true);
clsparse::vector<T> p(control, N, 0, CL_MEM_READ_WRITE, true);
clsparse::scalar<T> one(control, 1, CL_MEM_READ_ONLY, true);
clsparse::scalar<T> zero(control, 0, CL_MEM_READ_ONLY, true);
// y = A*x
status = csrmv<T>(one, pA, x, zero, y, control);
CLSPARSE_V(status, "csrmv Failed");
//r = b - y
status = r.sub(b, y, control);
//status = elementwise_transform<T, EW_MINUS>(r, b, y, control);
CLSPARSE_V(status, "b - y Failed");
clsparse::scalar<T> norm_r(control, 0, CL_MEM_WRITE_ONLY, false);
status = Norm1<T>(norm_r, r, control);
CLSPARSE_V(status, "norm r Failed");
//T residuum = 0;
clsparse::scalar<T> residuum(control, 0, CL_MEM_WRITE_ONLY, false);
//residuum = norm_r[0] / h_norm_b;
residuum.div(norm_r, norm_b, control);
solverControl->initialResidual = residuum[0];
#ifndef NDEBUG
std::cout << "initial residuum = "
<< solverControl->initialResidual << std::endl;
#endif
if (solverControl->finished(solverControl->initialResidual))
{
solverControl->nIters = 0;
return clsparseSuccess;
}
//apply preconditioner z = M*r
M(r, z, control);
//copy inital z to p
p = z;
//rz = <r, z>, here actually should be conjugate(r)) but we do not support complex type.
clsparse::scalar<T> rz(control, 0, CL_MEM_WRITE_ONLY, false);
status = dot<T>(rz, r, z, control);
CLSPARSE_V(status, "<r, z> Failed");
int iteration = 0;
bool converged = false;
clsparse::scalar<T> alpha (control, 0, CL_MEM_READ_WRITE, false);
clsparse::scalar<T> beta (control, 0, CL_MEM_READ_WRITE, false);
//yp buffer for inner product of y and p vectors;
clsparse::scalar<T> yp(control, 0, CL_MEM_WRITE_ONLY, false);
clsparse::scalar<T> rz_old(control, 0, CL_MEM_WRITE_ONLY, false);
while(!converged)
{
solverControl->nIters = iteration;
//y = A*p
status = csrmv<T>(one, pA, p, zero, y, control);
CLSPARSE_V(status, "csrmv Failed");
status = dot<T>(yp, y, p, control);
CLSPARSE_V(status, "<y,p> Failed");
// alpha = <r,z> / <y,p>
//alpha[0] = rz[0] / yp[0];
alpha.div(rz, yp, control);
#ifndef NDEBUG
std::cout << "alpha = " << alpha[0] << std::endl;
#endif
//x = x + alpha*p
status = axpy<T>(x, alpha, p, x, control);
CLSPARSE_V(status, "x = x + alpha * p Failed");
//r = r - alpha * y;
status = axpy<T, EW_MINUS>(r, alpha, y, r, control);
CLSPARSE_V(status, "r = r - alpha * y Failed");
//apply preconditioner z = M*r
M(r, z, control);
//store old value of rz
//improve that by move or swap
rz_old = rz;
//rz = <r,z>
status = dot<T>(rz, r, z, control);
CLSPARSE_V(status, "<r,z> Failed");
// beta = <r^(i), r^(i)>/<r^(i-1),r^(i-1)> // i: iteration index;
// beta is ratio of dot product in current iteration compared
//beta[0] = rz[0] / rz_old[0];
beta.div(rz, rz_old, control);
#ifndef NDEBUG
std::cout << "beta = " << beta[0] << std::endl;
#endif
//p = z + beta*p;
status = axpby<T>(p, one, z, beta, p, control );
CLSPARSE_V(status, "p = z + beta*p Failed");
//calculate norm of r
status = Norm1<T>(norm_r, r, control);
CLSPARSE_V(status, "norm r Failed");
//residuum = norm_r[0] / h_norm_b;
status = residuum.div(norm_r, norm_b, control);
CLSPARSE_V(status, "residuum");
iteration++;
converged = solverControl->finished(residuum[0]);
solverControl->print();
}
return clsparseSuccess;
}
int main(int argc, char* argv[])
{
//
// Get a default output stream from the Teuchos::VerboseObjectBase
//
Teuchos::RCP<Teuchos::FancyOStream>
out = Teuchos::VerboseObjectBase::getDefaultOStream();
//
// Set the parameters for the Belos LOWS Factory and create a parameter list.
//
int blockSize = 2;
int maxIterations = 400;
int maxRestarts = 25;
int gmresKrylovLength = 25;
int outputFrequency = 1;
bool outputMaxResOnly = true;
double maxResid = 1e-6;
Teuchos::RCP<Teuchos::ParameterList>
belosLOWSFPL = Teuchos::rcp( new Teuchos::ParameterList() );
belosLOWSFPL->set("Solver Type","Block GMRES");
Teuchos::ParameterList& belosLOWSFPL_solver =
belosLOWSFPL->sublist("Solver Types");
Teuchos::ParameterList& belosLOWSFPL_gmres =
belosLOWSFPL_solver.sublist("Block GMRES");
belosLOWSFPL_gmres.set("Maximum Iterations",int(maxIterations));
belosLOWSFPL_gmres.set("Convergence Tolerance",double(maxResid));
belosLOWSFPL_gmres.set("Maximum Restarts",int(maxRestarts));
belosLOWSFPL_gmres.set("Block Size",int(blockSize));
belosLOWSFPL_gmres.set("Num Blocks",int(gmresKrylovLength));
belosLOWSFPL_gmres.set("Output Frequency",int(outputFrequency));
belosLOWSFPL_gmres.set("Show Maximum Residual Norm Only",bool(outputMaxResOnly));
#ifdef HAVE_BELOS_IFPACK
//
// Set the parameters for the Ifpack Preconditioner Factory and create parameter list
//
Teuchos::ParameterList &ifpackPFSL = belosLOWSFPL->sublist("IfpackPreconditionerFactory");
ifpackPFSL.set("Overlap",int(2));
ifpackPFSL.set("Prec Type","ILUT");
#endif
// Whether the linear solver succeeded.
// (this will be set during the residual check at the end)
bool success = true;
// Number of random right-hand sides we will be solving for.
int numRhs = 5;
// Name of input matrix file
std::string matrixFile = "orsirr1.hb";
// Read in the matrix file (can be *.mtx, *.hb, etc.)
Epetra_SerialComm comm;
Teuchos::RCP<Epetra_CrsMatrix> epetra_A;
EpetraExt::readEpetraLinearSystem( matrixFile, comm, &epetra_A );
// Create a Thyra linear operator (A) using the Epetra_CrsMatrix (epetra_A).
Teuchos::RCP<const Thyra::LinearOpBase<double> >
A = Thyra::epetraLinearOp(epetra_A);
// Get the domain space for the Thyra linear operator
Teuchos::RCP<const Thyra::VectorSpaceBase<double> > domain = A->domain();
// Create the Belos LOWS factory.
Teuchos::RCP<Thyra::LinearOpWithSolveFactoryBase<double> >
belosLOWSFactory = Teuchos::rcp(new Thyra::BelosLinearOpWithSolveFactory<double>());
#ifdef HAVE_BELOS_IFPACK
// Set the preconditioner factory for the LOWS factory.
belosLOWSFactory->setPreconditionerFactory(
Teuchos::rcp(new Thyra::IfpackPreconditionerFactory())
,"IfpackPreconditionerFactory"
);
#endif
// Set the parameter list to specify the behavior of the factory.
belosLOWSFactory->setParameterList( belosLOWSFPL );
// Set the output stream and the verbosity level (prints to std::cout by defualt)
belosLOWSFactory->setVerbLevel(Teuchos::VERB_LOW);
// Create a BelosLinearOpWithSolve object from the Belos LOWS factory.
Teuchos::RCP<Thyra::LinearOpWithSolveBase<double> >
nsA = belosLOWSFactory->createOp();
// Initialize the BelosLinearOpWithSolve object with the Thyra linear operator.
Thyra::initializeOp<double>( *belosLOWSFactory, A, &*nsA );
// Create a right-hand side with numRhs vectors in it and randomize it.
Teuchos::RCP< Thyra::MultiVectorBase<double> >
b = Thyra::createMembers(domain, numRhs);
Thyra::seed_randomize<double>(0);
Thyra::randomize(-1.0, 1.0, &*b);
// Create an initial std::vector with numRhs vectors in it and initialize it to zero.
Teuchos::RCP< Thyra::MultiVectorBase<double> >
x = Thyra::createMembers(domain, numRhs);
Thyra::assign(&*x, 0.0);
// Perform solve using the linear operator to get the approximate solution of Ax=b,
// where b is the right-hand side and x is the left-hand side.
Thyra::SolveStatus<double> solveStatus;
solveStatus = Thyra::solve( *nsA, Thyra::NONCONJ_ELE, *b, &*x );
// Print out status of solve.
*out << "\nBelos LOWS Status: "<< solveStatus << std::endl;
//
// Compute residual and double check convergence.
//
std::vector<double> norm_b(numRhs), norm_res(numRhs);
Teuchos::RCP< Thyra::MultiVectorBase<double> >
y = Thyra::createMembers(domain, numRhs);
// Compute the column norms of the right-hand side b.
Thyra::norms_2( *b, &norm_b[0] );
// Compute y=A*x, where x is the solution from the linear solver.
A->apply( Thyra::NONCONJ_ELE, *x, &*y );
// Compute A*x-b = y-b
Thyra::update( -1.0, *b, &*y );
// Compute the column norms of A*x-b.
Thyra::norms_2( *y, &norm_res[0] );
// Print out the final relative residual norms.
double rel_res = 0.0;
*out << "Final relative residual norms" << std::endl;
for (int i=0; i<numRhs; ++i) {
rel_res = norm_res[i]/norm_b[i];
if (rel_res > maxResid)
success = false;
*out << "RHS " << i+1 << " : "
<< std::setw(16) << std::right << rel_res << std::endl;
}
return ( success ? 0 : 1 );
}
int main(int argc, char* argv[])
{
//
// Get a default output stream from the Teuchos::VerboseObjectBase
//
Teuchos::RCP<Teuchos::FancyOStream>
out = Teuchos::VerboseObjectBase::getDefaultOStream();
Teuchos::GlobalMPISession mpiSession(&argc,&argv);
#ifdef HAVE_COMPLEX
typedef std::complex<double> ST; // Scalar-type typedef
#elif HAVE_COMPLEX_H
typedef std::complex<double> ST; // Scalar-type typedef
#else
typedef double ST; // Scalar-type typedef
#endif
typedef Teuchos::ScalarTraits<ST>::magnitudeType MT; // Magnitude-type typedef
typedef int OT; // Ordinal-type typedef
ST one = Teuchos::ScalarTraits<ST>::one();
ST zero = Teuchos::ScalarTraits<ST>::zero();
#ifdef HAVE_MPI
MPI_Comm mpiComm = MPI_COMM_WORLD;
const Tpetra::MpiPlatform<OT,OT> ordinalPlatform(mpiComm);
const Tpetra::MpiPlatform<OT,ST> scalarPlatform(mpiComm);
#else
const Tpetra::SerialPlatform<OT,OT> ordinalPlatform;
const Tpetra::SerialPlatform<OT,ST> scalarPlatform;
#endif
//
// Get the data from the HB file
//
// Name of input matrix file
std::string matrixFile = "mhd1280b.cua";
int info=0;
int dim,dim2,nnz;
MT *dvals;
int *colptr,*rowind;
ST *cvals;
nnz = -1;
info = readHB_newmat_double(matrixFile.c_str(),&dim,&dim2,&nnz,
&colptr,&rowind,&dvals);
if (info == 0 || nnz < 0) {
*out << "Error reading '" << matrixFile << "'" << std::endl;
}
#ifdef HAVE_MPI
MPI_Finalize();
#endif
// Convert interleaved doubles to std::complex values
cvals = new ST[nnz];
for (int ii=0; ii<nnz; ii++) {
cvals[ii] = ST(dvals[ii*2],dvals[ii*2+1]);
}
// Declare global dimension of the linear operator
OT globalDim = dim;
// Create the element space and std::vector space
const Tpetra::ElementSpace<OT> elementSpace(globalDim,0,ordinalPlatform);
const Tpetra::VectorSpace<OT,ST> vectorSpace(elementSpace,scalarPlatform);
// Create my implementation of a Tpetra::Operator
RCP<Tpetra::Operator<OT,ST> >
tpetra_A = rcp( new MyOperator<OT,ST>(vectorSpace,dim,colptr,nnz,rowind,cvals) );
// Create a Thyra linear operator (A) using the Tpetra::CisMatrix (tpetra_A).
RCP<Thyra::LinearOpBase<ST> >
A = Teuchos::rcp( new Thyra::TpetraLinearOp<OT,ST>(tpetra_A) );
//
// Set the parameters for the Belos LOWS Factory and create a parameter list.
//
int blockSize = 1;
int maxIterations = globalDim;
int maxRestarts = 15;
int gmresKrylovLength = 50;
int outputFrequency = 100;
bool outputMaxResOnly = true;
MT maxResid = 1e-5;
Teuchos::RCP<Teuchos::ParameterList>
belosLOWSFPL = Teuchos::rcp( new Teuchos::ParameterList() );
belosLOWSFPL->set("Solver Type","Block GMRES");
Teuchos::ParameterList& belosLOWSFPL_solver =
belosLOWSFPL->sublist("Solver Types");
Teuchos::ParameterList& belosLOWSFPL_gmres =
belosLOWSFPL_solver.sublist("Block GMRES");
belosLOWSFPL_gmres.set("Maximum Iterations",int(maxIterations));
belosLOWSFPL_gmres.set("Convergence Tolerance",MT(maxResid));
belosLOWSFPL_gmres.set("Maximum Restarts",int(maxRestarts));
belosLOWSFPL_gmres.set("Block Size",int(blockSize));
belosLOWSFPL_gmres.set("Num Blocks",int(gmresKrylovLength));
belosLOWSFPL_gmres.set("Output Frequency",int(outputFrequency));
belosLOWSFPL_gmres.set("Show Maximum Residual Norm Only",bool(outputMaxResOnly));
// Whether the linear solver succeeded.
// (this will be set during the residual check at the end)
bool success = true;
// Number of random right-hand sides we will be solving for.
int numRhs = 1;
// Get the domain space for the Thyra linear operator
Teuchos::RCP<const Thyra::VectorSpaceBase<ST> > domain = A->domain();
// Create the Belos LOWS factory.
Teuchos::RCP<Thyra::LinearOpWithSolveFactoryBase<ST> >
belosLOWSFactory = Teuchos::rcp(new Thyra::BelosLinearOpWithSolveFactory<ST>());
// Set the parameter list to specify the behavior of the factory.
belosLOWSFactory->setParameterList( belosLOWSFPL );
// Set the output stream and the verbosity level (prints to std::cout by defualt)
// NOTE: Set to VERB_NONE for no output from the solver.
belosLOWSFactory->setVerbLevel(Teuchos::VERB_LOW);
// Create a BelosLinearOpWithSolve object from the Belos LOWS factory.
Teuchos::RCP<Thyra::LinearOpWithSolveBase<ST> >
nsA = belosLOWSFactory->createOp();
// Initialize the BelosLinearOpWithSolve object with the Thyra linear operator.
Thyra::initializeOp<ST>( *belosLOWSFactory, A, &*nsA );
// Create a right-hand side with numRhs vectors in it.
Teuchos::RCP< Thyra::MultiVectorBase<ST> >
b = Thyra::createMembers(domain, numRhs);
// Create an initial std::vector with numRhs vectors in it and initialize it to one.
Teuchos::RCP< Thyra::MultiVectorBase<ST> >
x = Thyra::createMembers(domain, numRhs);
Thyra::assign(&*x, one);
// Initialize the right-hand side so that the solution is a std::vector of ones.
A->apply( Thyra::NONCONJ_ELE, *x, &*b );
Thyra::assign(&*x, zero);
// Perform solve using the linear operator to get the approximate solution of Ax=b,
// where b is the right-hand side and x is the left-hand side.
Thyra::SolveStatus<ST> solveStatus;
solveStatus = Thyra::solve( *nsA, Thyra::NONCONJ_ELE, *b, &*x );
// Print out status of solve.
*out << "\nBelos LOWS Status: "<< solveStatus << std::endl;
//
// Compute residual and ST check convergence.
//
std::vector<MT> norm_b(numRhs), norm_res(numRhs);
Teuchos::RCP< Thyra::MultiVectorBase<ST> >
y = Thyra::createMembers(domain, numRhs);
// Compute the column norms of the right-hand side b.
Thyra::norms_2( *b, &norm_b[0] );
// Compute y=A*x, where x is the solution from the linear solver.
A->apply( Thyra::NONCONJ_ELE, *x, &*y );
// Compute A*x-b = y-b
Thyra::update( -one, *b, &*y );
// Compute the column norms of A*x-b.
Thyra::norms_2( *y, &norm_res[0] );
// Print out the final relative residual norms.
MT rel_res = 0.0;
*out << "Final relative residual norms" << std::endl;
for (int i=0; i<numRhs; ++i) {
rel_res = norm_res[i]/norm_b[i];
if (rel_res > maxResid)
success = false;
*out << "RHS " << i+1 << " : "
<< std::setw(16) << std::right << rel_res << std::endl;
}
return ( success ? 0 : 1 );
}
