/** Set up the internal matrices used to accumulate image statistics,
* and initialize the batch size.
*/
void cv_mean_extractor::set(const unsigned int width, const unsigned int height,
const unsigned int n_ch, const unsigned int batch_sz) {
if (!m_sum.empty() || (width == 0u) || (height == 0u) || (n_ch == 0u) || (batch_sz == 0u)) {
std::stringstream err;
err << __FILE__ << " " << __LINE__ << " :: cv_mean_extractor: either using an invalid "
<< "parameter or attempting to reconfigure";
throw lbann_exception(err.str());
}
m_batch_size = batch_sz;
create_matrices(width, height, n_ch);
reset();
}
/**
* If the size or the number of channels of the given image is different
* from what is expected, fails.
*/
bool cv_mean_extractor::determine_transform(const cv::Mat& image) {
m_enabled = false;
_LBANN_SILENT_EXCEPTION(image.empty(), "", false);
// If it has not been configured (other than batch size), do it here
if (m_sum.empty()) {
create_matrices(image.cols, image.rows, image.channels());
reset();
m_enabled = true;
} else {
m_enabled = check_if_cv_Mat_has_same_shape(image, m_avg);
}
return m_enabled;
}
/* Find the DBBD form */
int shylu_symbolic_factor
(
Epetra_CrsMatrix *A, // i/p: A matrix
shylu_symbolic *ssym, // symbolic structure
shylu_data *data, // numeric structure, TODO: Required ?
shylu_config *config // i/p: library configuration
)
{
#ifdef TIMING_OUTPUT
Teuchos::Time symtime("symbolic time");
symtime.start();
#endif
int myPID = A->Comm().MyPID();
int n = A->NumGlobalRows();
int Dnr;
int Snr;
int *DRowElems;
int *SRowElems;
int sym = config->sym;
checkMaps(A);
// Get column map
Epetra_Map AColMap = A->ColMap();
int ncols = AColMap.NumMyElements();
int *cols = AColMap.MyGlobalElements();
// Get row map
Epetra_Map ARowMap = A->RowMap();
int nrows = ARowMap.NumMyElements();
int *rows = ARowMap.MyGlobalElements();
// Find all columns in this proc
int *gvals = new int[n]; // vector of size n, not ncols !
// gvals[local cols] = 1, gvals[shared cols] > 1.
int SNumGlobalCols;
findLocalColumns(A, gvals, SNumGlobalCols);
// See if you can shrink the separator by assigning more rows/columns to
// the block diagonals
// TODO: This is because of a bug in coloring remove the if once that is
// fixed
//if (config->schurApproxMethod == 2)
if (config->sep_type == 2)
findNarrowSeparator(A, gvals);
// 3. Assemble diagonal block and the border in convenient form [
/* In each processor, we have (in a permuted form)
* | D_i C_i |
* | R_i S_i |
* D_i - diagonal block, C_i - Column Separator, R_i - Row separator
* S_i - A22 block corresponding to Schur complement part of A
* Assemble all four blocks in local matrices. */
ostringstream ssmsg1;
ssmsg1 << "PID =" << myPID << " ";
string msg = ssmsg1.str();
ssmsg1.clear(); ssmsg1.str("");
// Find #cols in each block
int Dnc = 0; // #cols in diagonal block
int Snc = 0; // #cols in the col. separator
/* Looping on cols will work only for wide separator
* as for narrow sep there will be some sep cols with gvals[col] ==1
* */
/*for (int i = 0; i < ncols ; i++)
{
if (gvals[cols[i]] == 1)
Dnc++;
else
Snc++;
}
// Find #rows in each block
Dnr = Dnc; // #rows in square diagonal block
Snr = nrows - Dnr; // #rows in the row separator*/
// Find #rows in each block
Dnr = 0;
Snr = 0;
for (int i = 0; i < nrows ; i++)
{
if (gvals[rows[i]] == 1)
Dnr++;
else
Snr++;
}
Dnc = Dnr;
// TODO: Snc is no longer useful, should remove it
for (int i = 0; i < ncols ; i++)
{
if (gvals[cols[i]] != 1)
Snc++;
}
assert(Snc >= 0);
// TODO : The above assignment may not be correct in the unsymetric case
////config->dm.print(2, msg + " Mycols=");
cout << msg << " Mycols="<< ncols << "Myrows ="<< nrows << endl;
cout << msg << " #rows and #cols in diagonal blk ="<< Dnr << endl;
cout << msg << " #columns in S ="<< Snc << endl;
cout << msg << " #rows in S ="<< Snr << endl;
ostringstream pidstr;
pidstr << myPID ;
// Create a row map for the D and S blocks [
DRowElems = new int[Dnr];
SRowElems = new int[Snr];
int gid;
// Assemble row ids in two arrays (for D and R blocks)
if (sym)
{
findBlockElems(A, nrows, rows, gvals, Dnr, DRowElems, Snr, SRowElems,
"D"+pidstr.str()+"Rows", "S"+pidstr.str()+"Rows", false) ;
}
else
{
// SRowElems are not known until factorization, TODO
assert(0 == 1);
}
data->Dnr = Dnr;
data->Snr = Snr;
data->Dnc = Dnc;
data->DRowElems = DRowElems;
data->SRowElems = SRowElems;
// Create a column map for the D and S blocks [
int *DColElems = new int[Dnc]; // Elems in column map of D
int *SColElems = new int[Snc]; // Elems in column map of C TODO: Unused
// Assemble column ids in two arrays (for D and C blocks)
findBlockElems(A, ncols, cols, gvals, Dnc, DColElems, Snc, SColElems,
"D"+pidstr.str()+"Cols", "S"+pidstr.str()+"Cols", true) ;
data->DColElems = DColElems;
data->gvals = gvals;
for (int i = 0; i < Snr; i++)
{
// Epetra guarentees columns corresponding to local rows will be first
// in the column map.
assert(SRowElems[i] == SColElems[i]);
}
// ]
/*--Create the Epetra Matrices with the maps (does not insert values) --- */
create_matrices(A, ssym, data, config);
/*--Extract the Epetra Matrices and call fillComplete --- */
extract_matrices(A, ssym, data, config, true);
delete[] SColElems;
Amesos Factory;
const char* SolverType = config->diagonalBlockSolver.c_str();
bool IsAvailable = Factory.Query(SolverType);
assert(IsAvailable == true);
Teuchos::RCP<Epetra_LinearProblem> LP = Teuchos::RCP<Epetra_LinearProblem>
(new Epetra_LinearProblem());
LP->SetOperator((ssym->D).getRawPtr());
//LP->SetOperator((ssym->DT).getRawPtr()); // for transpose
// Create temp vectors
ssym->Dlhs = Teuchos::RCP<Epetra_MultiVector>
(new Epetra_MultiVector(ssym->D->RowMap(), 16));
ssym->Drhs = Teuchos::RCP<Epetra_MultiVector>
(new Epetra_MultiVector(ssym->D->RowMap(), 16));
ssym->Gvec = Teuchos::RCP<Epetra_MultiVector>
(new Epetra_MultiVector(ssym->G->RowMap(), 16));
LP->SetRHS(ssym->Drhs.getRawPtr());
LP->SetLHS(ssym->Dlhs.getRawPtr());
ssym->ReIdx_LP = Teuchos::RCP<
EpetraExt::ViewTransform<Epetra_LinearProblem> >
(new EpetraExt::LinearProblem_Reindex2(0));
ssym->LP = Teuchos::RCP<Epetra_LinearProblem>(&((*(ssym->ReIdx_LP))(*LP)),
false);
Teuchos::RCP<Amesos_BaseSolver> Solver = Teuchos::RCP<Amesos_BaseSolver>
(Factory.Create(SolverType, *(ssym->LP)));
//config->dm.print(5, "Created the diagonal solver");
#ifdef TIMING_OUTPUT
Teuchos::Time ftime("setup time");
ftime.start();
#endif
//Solver->SetUseTranspose(true); // for transpose
Teuchos::ParameterList aList;
aList.set("TrustMe", true);
Solver->SetParameters(aList);
Solver->SymbolicFactorization();
//config->dm.print(3, "Symbolic Factorization done");
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Symbolic Factorization Time" << ftime.totalElapsedTime() << endl;
ftime.reset();
#endif
ssym->OrigLP = LP;
//ssym->LP = LP;
ssym->Solver = Solver;
if (config->schurApproxMethod == 1)
{
Teuchos::ParameterList pList;
Teuchos::RCP<Isorropia::Epetra::Prober> prober =
Teuchos::RCP<Isorropia::Epetra::Prober> (new
Isorropia::Epetra::Prober((ssym->Sg).getRawPtr(),
pList, false));
//config->dm.print(3, "Doing Coloring");
#ifdef TIMING_OUTPUT
ftime.start();
#endif
prober->color();
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Time to color" << ftime.totalElapsedTime() << endl;
ftime.reset();
ftime.start();
#endif
ssym->prober = prober;
}
#ifdef TIMING_OUTPUT
symtime.stop();
cout << "Symbolic Time" << symtime.totalElapsedTime() << endl;
symtime.reset();
#endif
}