502 lines
16 KiB
C++
502 lines
16 KiB
C++
/*
|
|
* =====================================================================================
|
|
*
|
|
* Filename: sparse_matrix_impl.h
|
|
*
|
|
* Description:
|
|
*
|
|
* Version: 1.0
|
|
* Created: 12/02/2007 10:18:02 AM EST
|
|
* Revision: none
|
|
* Compiler: gcc
|
|
*
|
|
* Author: Nikolaos Vasiloglou (NV), nvasil@ieee.org
|
|
* Company: Georgia Tech Fastlab-ESP Lab
|
|
*
|
|
* =====================================================================================
|
|
*
|
|
*/
|
|
|
|
SparseMatrix::SparseMatrix() {
|
|
map_ = NULL;
|
|
issymmetric_=false;
|
|
}
|
|
|
|
SparseMatrix::SparseMatrix(const index_t num_of_rows,
|
|
const index_t num_of_columns,
|
|
const index_t nnz_per_row) {
|
|
Init(num_of_rows, num_of_columns, nnz_per_row);
|
|
}
|
|
void SparseMatrix::Init(const index_t num_of_rows,
|
|
const index_t num_of_columns,
|
|
const index_t nnz_per_row) {
|
|
issymmetric_ = false;
|
|
if likely(num_of_rows < num_of_columns) {
|
|
FATAL("Num of rows %i should be greater than the num of columns %i\n",
|
|
num_of_rows, num_of_columns);
|
|
}
|
|
num_of_rows_=num_of_rows;
|
|
num_of_columns_=num_of_columns;
|
|
dimension_ = num_of_rows_;
|
|
if (num_of_rows_ < num_of_columns_) {
|
|
FATAL("Num of rows %i is less than the number or columns %i",
|
|
num_of_rows_, num_of_columns_);
|
|
}
|
|
map_ = new Epetra_Map(num_of_rows_, 0, comm_);
|
|
matrix_ = Teuchos::rcp(
|
|
new Epetra_CrsMatrix((Epetra_DataAccess)0, *map_, nnz_per_row));
|
|
StartLoadingRows();
|
|
}
|
|
|
|
void SparseMatrix::Init(const Epetra_CrsMatrix &other) {
|
|
issymmetric_ = false;
|
|
matrix_ = Teuchos::rcp(new Epetra_CrsMatrix(other));
|
|
map_ = new Epetra_Map(matrix_->RowMap());
|
|
dimension_ = other.NumGlobalRows();
|
|
num_of_rows_=dimension_;
|
|
num_of_columns_=dimension_;
|
|
}
|
|
|
|
void SparseMatrix::Init(index_t num_of_rows,
|
|
index_t num_of_columns,
|
|
index_t *nnz_per_row) {
|
|
issymmetric_ = false;
|
|
num_of_rows_=num_of_rows;
|
|
num_of_columns_=num_of_columns;
|
|
dimension_ = num_of_rows;
|
|
if (num_of_rows_ < num_of_columns_) {
|
|
FATAL("Num of rows %i is less than the number or columns %i",
|
|
num_of_rows_, num_of_columns_);
|
|
}
|
|
map_ = new Epetra_Map(num_of_rows_, 0, comm_);
|
|
matrix_ = Teuchos::rcp(
|
|
new Epetra_CrsMatrix((Epetra_DataAccess)0 , *map_, nnz_per_row));
|
|
StartLoadingRows();
|
|
}
|
|
|
|
|
|
void SparseMatrix::Init(const std::vector<index_t> &rows,
|
|
const std::vector<index_t> &columns,
|
|
const Vector &values,
|
|
index_t nnz_per_row,
|
|
index_t dimension) {
|
|
issymmetric_ = false;
|
|
if (nnz_per_row > 0 && dimension > 0) {
|
|
Init(dimension, dimension, nnz_per_row);
|
|
} else {
|
|
num_of_rows_ = 0;
|
|
num_of_columns_=0;
|
|
std::map<index_t, index_t> frequencies;
|
|
for(index_t i=0; i<(index_t)rows.size(); i++) {
|
|
frequencies[rows[i]]++;
|
|
frequencies[columns[i]]++;
|
|
if (rows[i]>num_of_rows_) {
|
|
num_of_rows_ = rows[i];
|
|
}
|
|
if (columns[i]>num_of_columns_) {
|
|
num_of_columns_ = columns[i];
|
|
}
|
|
}
|
|
num_of_columns_++;
|
|
num_of_rows_++;
|
|
if (num_of_rows_ < num_of_columns_) {
|
|
FATAL("At this point we only support rows (%i) >= columns (%i)\n",
|
|
num_of_rows_, num_of_columns_);
|
|
}
|
|
dimension_ = num_of_rows_;
|
|
if ((index_t)frequencies.size()!=num_of_rows_) {
|
|
NONFATAL("Some of the rows are zeros only!");
|
|
}
|
|
index_t *nnz= new index_t[dimension_];
|
|
for(index_t i=0; i<num_of_rows_; i++) {
|
|
nnz[i]=frequencies[i];
|
|
}
|
|
Init(num_of_rows_, num_of_columns_, nnz);
|
|
//delete []nnz;
|
|
}
|
|
Load(rows, columns, values);
|
|
}
|
|
|
|
void SparseMatrix::Init(const std::vector<index_t> &rows,
|
|
const std::vector<index_t> &columns,
|
|
const std::vector<double> &values,
|
|
index_t nnz_per_row,
|
|
index_t dimension) {
|
|
issymmetric_ = false;
|
|
Vector temp;
|
|
temp.Alias((double *)&values[0], values.size());
|
|
Init(rows, columns, temp, nnz_per_row, dimension);
|
|
|
|
}
|
|
|
|
void SparseMatrix::Init(std::string filename) {
|
|
issymmetric_ = false;
|
|
FILE *fp = fopen(filename.c_str(), "r");
|
|
if (fp == NULL) {
|
|
FATAL("Cannot open %s, error: %s",
|
|
filename.c_str(),
|
|
strerror(errno));
|
|
}
|
|
std::vector<index_t> rows;
|
|
std::vector<index_t> cols;
|
|
std::vector<double> vals;
|
|
while (!feof(fp)) {
|
|
index_t r, c;
|
|
double v;
|
|
fscanf(fp,"%i %i %lg\n", &r, &c, &v);
|
|
rows.push_back(r);
|
|
cols.push_back(c);
|
|
vals.push_back(v);
|
|
}
|
|
fclose(fp);
|
|
/*for(index_t i=0; i< (index_t)rows.size(); i++) {
|
|
printf("%i %i %lg\n", rows[i], cols[i], vals[i]);
|
|
}*/
|
|
Init(rows, cols, vals, -1, -1);
|
|
}
|
|
|
|
void SparseMatrix::Destruct() {
|
|
if (map_!=NULL) {
|
|
delete map_;
|
|
map_=NULL;
|
|
}
|
|
}
|
|
|
|
void SparseMatrix::Copy(const SparseMatrix &other) {
|
|
num_of_rows_ = other.num_of_rows_;
|
|
num_of_columns_ = other.num_of_columns_;
|
|
dimension_ = other.dimension_;
|
|
map_ = new Epetra_Map(num_of_rows_, 0, comm_);
|
|
issymmetric_ = other.issymmetric_;
|
|
if (other.matrix_->Filled()==false) {
|
|
matrix_ = Teuchos::rcp(new Epetra_CrsMatrix(Epetra_DataAccess(0), *map_, 10));
|
|
this->StartLoadingRows();
|
|
for(index_t r=0; r<num_of_rows_; r++){
|
|
index_t global_row = other.my_global_elements_[r];
|
|
index_t num_of_entries;
|
|
double *values;
|
|
index_t *indices;
|
|
other.matrix_->ExtractGlobalRowView(global_row, num_of_entries, values, indices);
|
|
this->LoadRow(r, num_of_entries, indices, values);
|
|
}
|
|
} else {
|
|
matrix_ = Teuchos::rcp(new Epetra_CrsMatrix(*(other.matrix_.get())));
|
|
}
|
|
}
|
|
|
|
void SparseMatrix::SetDiagonal(const Vector &vector) {
|
|
if (unlikely(vector.length()!=dimension_)) {
|
|
FATAL("Vector should have the same dimension with the matrix!\n");
|
|
}
|
|
for(index_t i=0; i<dimension_; i++) {
|
|
set(i,i, vector[i]);
|
|
}
|
|
}
|
|
|
|
void SparseMatrix::StartLoadingRows() {
|
|
my_global_elements_ = map_->MyGlobalElements();
|
|
}
|
|
|
|
void SparseMatrix::LoadRow(index_t row,
|
|
std::vector<index_t> &columns,
|
|
Vector &values) {
|
|
DEBUG_ASSERT(values.length() == (index_t)columns.size());
|
|
matrix_->InsertGlobalValues(my_global_elements_[row],
|
|
values.length(),
|
|
values.ptr(),
|
|
&columns[0]);
|
|
|
|
}
|
|
|
|
void SparseMatrix::LoadRow(index_t row,
|
|
index_t *columns,
|
|
Vector &values) {
|
|
matrix_->InsertGlobalValues(my_global_elements_[row],
|
|
values.length(),
|
|
values.ptr(),
|
|
&columns[0]);
|
|
|
|
}
|
|
|
|
void SparseMatrix::LoadRow(index_t row,
|
|
std::vector<index_t> &columns,
|
|
std::vector<double> &values) {
|
|
matrix_->InsertGlobalValues(my_global_elements_[row],
|
|
values.size(),
|
|
&values[0],
|
|
&columns[0]);
|
|
|
|
}
|
|
|
|
void SparseMatrix::LoadRow(index_t row,
|
|
index_t num,
|
|
index_t *columns,
|
|
double *values) {
|
|
matrix_->InsertGlobalValues(my_global_elements_[row],
|
|
num,
|
|
values,
|
|
columns);
|
|
|
|
}
|
|
|
|
void SparseMatrix::EndLoading() {
|
|
matrix_->FillComplete();
|
|
matrix_->OptimizeStorage();
|
|
}
|
|
|
|
void SparseMatrix::Load(const std::vector<index_t> &rows,
|
|
const std::vector<index_t> &columns,
|
|
const Vector &values) {
|
|
DEBUG_ASSERT(rows.size() ==columns.size());
|
|
DEBUG_ASSERT((index_t)columns.size() == values.length());
|
|
my_global_elements_ = map_->MyGlobalElements();
|
|
index_t i=0;
|
|
index_t cur_row = rows[i];
|
|
index_t prev_row= rows[i];
|
|
std::vector<index_t> indices;
|
|
std::vector<double> row_values;
|
|
while (true) {
|
|
indices.clear();
|
|
row_values.clear();
|
|
while (likely((rows[cur_row]==rows[prev_row]) &&
|
|
(i < (index_t)rows.size()))) {
|
|
indices.push_back(columns[i]);
|
|
row_values.push_back(values[i]);
|
|
i++;
|
|
prev_row=i-1;
|
|
cur_row=i;
|
|
}
|
|
matrix_->InsertGlobalValues(my_global_elements_[rows[prev_row]],
|
|
row_values.size(),
|
|
&row_values[0],
|
|
&indices[0]);
|
|
prev_row=cur_row;
|
|
if (i >= (index_t)rows.size()) {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
|
|
double SparseMatrix::get(index_t r, index_t c) const {
|
|
DEBUG_BOUNDS(r, num_of_rows_);
|
|
DEBUG_BOUNDS(c, num_of_columns_);
|
|
index_t global_row = my_global_elements_[r];
|
|
index_t num_of_entries;
|
|
double *values;
|
|
index_t *indices;
|
|
matrix_->ExtractGlobalRowView(global_row, num_of_entries, values, indices);
|
|
index_t *pos = std::find(indices, indices+num_of_entries, c);
|
|
if (pos==indices+num_of_entries) {
|
|
return 0;
|
|
}
|
|
return values[(ptrdiff_t)(pos-indices)];
|
|
}
|
|
|
|
void SparseMatrix::set(index_t r, index_t c, double v) {
|
|
DEBUG_BOUNDS(r, num_of_rows_);
|
|
DEBUG_BOUNDS(c, num_of_columns_);
|
|
if (get(r,c)!=0) {
|
|
matrix_->InsertGlobalValues(my_global_elements_[r], 1, &v, &c);
|
|
} else {
|
|
matrix_->ReplaceGlobalValues(my_global_elements_[r], 1, &v, &c);
|
|
}
|
|
}
|
|
|
|
void SparseMatrix::MakeSymmetric() {
|
|
index_t num_of_entries;
|
|
double *values;
|
|
index_t *indices;
|
|
my_global_elements_ = map_->MyGlobalElements();
|
|
for(index_t i=0; i<dimension_; i++) {
|
|
index_t global_row = my_global_elements_[i];
|
|
matrix_->ExtractGlobalRowView(global_row, num_of_entries, values, indices);
|
|
for(index_t j=0; j<num_of_entries; j++) {
|
|
if (unlikely(get(i, indices[j])!=values[j])) {
|
|
set(i, indices[j], values[j]);
|
|
}
|
|
}
|
|
}
|
|
issymmetric_ = true;
|
|
}
|
|
|
|
void SparseMatrix::Eig(index_t num_of_eigvalues,
|
|
std::string eigtype,
|
|
Matrix *eigvectors,
|
|
std::vector<double> *real_eigvalues,
|
|
std::vector<double> *imag_eigvalues) {
|
|
if (unlikely(!matrix_->Filled())) {
|
|
FATAL("You have to call EndLoading before running eigenvalues otherwise "
|
|
"it will fail\n");
|
|
}
|
|
|
|
index_t block_size=10;
|
|
retry:
|
|
Teuchos::RCP<Epetra_MultiVector> ivec = Teuchos::rcp(new
|
|
Epetra_MultiVector(*map_,
|
|
block_size));
|
|
// Fill it with random numbers
|
|
ivec->Random();
|
|
// Setup the eigenproblem, with the matrix A and the initial vectors ivec
|
|
Teuchos::RCP<Anasazi::BasicEigenproblem<double,MV,OP> >problem =
|
|
Teuchos::rcp(new Anasazi::BasicEigenproblem<double,MV,OP>(matrix_, ivec));
|
|
|
|
// The 2-D laplacian is symmetric. Specify this in the eigenproblem.
|
|
if (issymmetric_ == true) {
|
|
problem->setHermitian(true);
|
|
} else {
|
|
problem->setHermitian(false);
|
|
}
|
|
|
|
// Specify the desired number of eigenvalues
|
|
problem->setNEV(num_of_eigvalues);
|
|
|
|
// Signal that we are done setting up the eigenvalue problem
|
|
bool ierr = problem->setProblem();
|
|
|
|
// Check the return from setProblem(). If this is true, there was an
|
|
// error. This probably means we did not specify enough information for
|
|
// the eigenproblem.
|
|
if unlikely(ierr == false) {
|
|
FATAL("Trilinos solver error, you probably didn't specify enough information "
|
|
"for the eigenvalue problem\n");
|
|
}
|
|
|
|
// Specify the verbosity level. Options include:
|
|
// Anasazi::Errors
|
|
// This option is always set
|
|
// Anasazi::Warnings
|
|
// Warnings (less severe than errors)
|
|
// Anasazi::IterationDetails
|
|
// Details at each iteration, such as the current eigenvalues
|
|
// Anasazi::OrthoDetails
|
|
// Details about orthogonality
|
|
// Anasazi::TimingDetails
|
|
// A summary of the timing info for the solve() routine
|
|
// Anasazi::FinalSummary
|
|
// A final summary
|
|
// Anasazi::Debug
|
|
// Debugging information
|
|
int verbosity = Anasazi::Warnings +
|
|
Anasazi::Errors +
|
|
Anasazi::FinalSummary +
|
|
// Anasazi::IterationDetails +
|
|
Anasazi::TimingDetails;
|
|
|
|
// Choose which eigenvalues to compute
|
|
// Choices are:
|
|
// LM - target the largest magnitude [default]
|
|
// SM - target the smallest magnitude
|
|
// LR - target the largest real
|
|
// SR - target the smallest real
|
|
// LI - target the largest imaginary
|
|
// SI - target the smallest imaginary
|
|
// Create the parameter list for the eigensolver
|
|
double tolerance=1.0e-7;
|
|
|
|
Teuchos::ParameterList my_pl;
|
|
my_pl.set( "Verbosity", verbosity);
|
|
my_pl.set( "Which", eigtype);
|
|
my_pl.set( "Block Size", block_size);
|
|
my_pl.set( "Num Blocks", 20);
|
|
my_pl.set( "Maximum Restarts", 2);
|
|
my_pl.set( "Convergence Tolerance", tolerance);
|
|
|
|
// Create the Block Krylov Schur solver
|
|
// This takes as inputs the eigenvalue problem and the solver parameters
|
|
Teuchos::RCP<Anasazi::BlockKrylovSchurSolMgr<double,MV,OP> > my_block_krylov_schur;
|
|
my_block_krylov_schur=Teuchos::rcp(new
|
|
Anasazi::BlockKrylovSchurSolMgr<double,MV,OP> (problem, my_pl));
|
|
|
|
// Solve the eigenvalue problem, and save the return code
|
|
Anasazi::ReturnType solver_return = my_block_krylov_schur->solve();
|
|
|
|
// Check return code of the solver: Unconverged, Failed, or OK
|
|
switch (solver_return) {
|
|
// UNCONVERGED
|
|
case Anasazi::Unconverged:
|
|
block_size*=2;
|
|
if (block_size<64) {
|
|
tolerance=tolerance*10;
|
|
NONFATAL("Didn't converge, increasing block_size to %i\n and retrying ",
|
|
(int)block_size);
|
|
goto retry;
|
|
}
|
|
FATAL("Anasazi::BlockKrylovSchur::solve() did not converge!\n");
|
|
return ;
|
|
// CONVERGED
|
|
case Anasazi::Converged:
|
|
NONFATAL("Anasazi::BlockKrylovSchur::solve() converged!\n");
|
|
}
|
|
// Get eigensolution struct
|
|
Anasazi::Eigensolution<double, Epetra_MultiVector> sol = problem->getSolution();
|
|
// Get the number of eigenpairs returned
|
|
int num_of_eigvals_returned = sol.numVecs;
|
|
if (num_of_eigvals_returned < num_of_eigvalues) {
|
|
NONFATAL("The solver returned less eigenvalues (%i) "
|
|
"than requested (%i)\n", num_of_eigvalues,
|
|
num_of_eigvals_returned);
|
|
}
|
|
|
|
// Get eigenvectors
|
|
eigvectors->Init(dimension_, num_of_eigvals_returned);
|
|
Teuchos::RCP<Epetra_MultiVector> evecs = sol.Evecs;
|
|
evecs->ExtractCopy(eigvectors->GetColumnPtr(0), dimension_);
|
|
|
|
// Get eigenvalues
|
|
std::vector<Anasazi::Value<double> > evals = sol.Evals;
|
|
real_eigvalues->resize(0);
|
|
for(index_t i=0; i<num_of_eigvals_returned; i++) {
|
|
real_eigvalues->push_back(evals[i].realpart);
|
|
if (issymmetric_ == false) {
|
|
imag_eigvalues->resize(num_of_eigvals_returned);
|
|
imag_eigvalues->assign(i, evals[i].imagpart);
|
|
}
|
|
}
|
|
|
|
// Test residuals
|
|
// Generate a (numev x numev) dense matrix for the eigenvalues...
|
|
// This matrix is automatically initialized to zero
|
|
Teuchos::SerialDenseMatrix<int, double> d(num_of_eigvalues,
|
|
num_of_eigvalues);
|
|
|
|
// Add the eigenvalues on the diagonals (only the real part since problem is Hermitian)
|
|
for (int i=0; i<num_of_eigvalues; i++) {
|
|
d(i,i) = evals[i].realpart;
|
|
}
|
|
|
|
// Generate a multivector for the product of the matrix and the eigenvectors
|
|
Epetra_MultiVector res(*map_, num_of_eigvalues);
|
|
|
|
// R = A*evecs
|
|
matrix_->Apply( *evecs, res);
|
|
|
|
// R -= evecs*D
|
|
// = A*evecs - evecs*D
|
|
MVT::MvTimesMatAddMv( -1.0, *evecs, d, 1.0, res);
|
|
|
|
// Compute the 2-norm of each vector in the MultiVector
|
|
// and store them to a std::vector<double>
|
|
std::vector<double> norm_res(num_of_eigvalues);
|
|
MVT::MvNorm(res, &norm_res);
|
|
}
|
|
|
|
void SparseMatrix::IncompleteCholesky(index_t level_fill,
|
|
double drop_tol,
|
|
SparseMatrix *u,
|
|
Vector *d,
|
|
double *condest) {
|
|
Ifpack_CrsIct *ict=NULL;
|
|
ict = new Ifpack_CrsIct(*matrix_, drop_tol, level_fill);
|
|
// Init values from A
|
|
ict->InitValues(*matrix_);
|
|
// compute the factors
|
|
if (unlikely(ict->Factor()<0)) {
|
|
NONFATAL("Cholesky factorization failed!\n");
|
|
};
|
|
// and now estimate the condition number
|
|
ict->Condest(false, *condest);
|
|
u->Init(ict->U());
|
|
ict->D().ExtractCopy(d->ptr());
|
|
}
|
|
|