void random_distribution_1D(
itype nrows, // Number of global matrix rows
Epetra_Comm &comm, // Epetra communicator to be used in maps
Epetra_Map **rowMap, // OUTPUT: pointer to row map to be created
long long offsetEpetra64
)
{
// Randomly assign matrix rows to processor's row Map.
int me = comm.MyPID();
int np = comm.NumProc();
vector<itype> myGlobalElements(1.2 * (nrows / np) + 1);
int nMyRows = 0;
srandom(1);
double denom = (double) RAND_MAX + 1.;
for (itype i = 0; i < nrows; i++) {
int p = (int) ((double) np * (double) random() / denom);
if (p == me) {
if (nMyRows >= myGlobalElements.size())
myGlobalElements.resize(1.5*myGlobalElements.size());
myGlobalElements[nMyRows] = i + offsetEpetra64;
nMyRows++;
}
}
*rowMap = new Epetra_Map(nrows, nMyRows, &myGlobalElements[0], 0, comm);
}
bool global_check_for_flag_on_proc_0(const char* flag,
int numargs,
char** strargs,
const Epetra_Comm& comm)
{
int mypid = comm.MyPID();
int numprocs = comm.NumProc();
int flag_found = 0;
if (mypid==0) {
for(int i=0; i<numargs; ++i) {
if (strargs[i]==0) continue;
if (strcmp(flag, strargs[i]) == 0) {
flag_found = 1;
break;
}
}
}
if (numprocs > 1) {
comm.Broadcast(&flag_found, 1, 0);
}
bool return_value = flag_found==1 ? true : false;
return( return_value );
}
int fevec6(Epetra_Comm& Comm, bool verbose)
{
int NumElements = 4;
Epetra_Map Map(NumElements, 0, Comm);
Epetra_FEVector x1(Map);
x1.PutScalar (0);
// let all processors set global entry 0 to 1
const int GID = 0;
const double value = 1;
x1.ReplaceGlobalValues(1, &GID, &value);
x1.GlobalAssemble (Insert);
if (Comm.MyPID()==0)
std::cout << "Entry " << GID << " after construct & set: "
<< x1[0][0] << std::endl;
x1.PutScalar(0);
// re-apply 1 to the vector, but only on the
// owning processor. should be enough to set
// the value (as non-local data in x1 should
// have been eliminated after calling
// GlobalAssemble).
if (Comm.MyPID()==0)
x1.ReplaceGlobalValues(1, &GID, &value);
x1.GlobalAssemble (Insert);
if (Comm.MyPID()==0) {
std::cout << "Entry " << GID << " after PutScalar & set: "
<< x1[0][0] << std::endl;
if (x1[0][0] != value) return -1;
}
return 0;
}
void ML_Read_Matrix_Dimensions(const char *filename, int *numGlobalRows, Epetra_Comm &Comm)
{
char line[35], token1[35], token2[35], token3[35], token4[35], token5[35];
int lineLength = 1025;
FILE *fid = fopen(filename,"r");
int N, NZ;
if(fgets(line, lineLength, fid)==0) {
if (fid!=0) fclose(fid);
ML_Exit(Comm.MyPID(),"error opening matrix file", EXIT_FAILURE);
}
if(sscanf(line, "%s %s %s %s %s", token1, token2, token3, token4, token5 )==0) {
if (fid!=0) fclose(fid);
ML_Exit(Comm.MyPID(),"error reading matrix file header", EXIT_FAILURE);
}
if (strcmp(token1, "%%MatrixMarket") || strcmp(token2, "matrix") ||
strcmp(token3, "coordinate") || strcmp(token4, "real") ||
strcmp(token5, "general"))
{
if (fid!=0) fclose(fid);
ML_Exit(Comm.MyPID(),"error reading matrix file header", EXIT_FAILURE);
}
// Next, strip off header lines (which start with "%")
do {
if(fgets(line, lineLength, fid)==0) {
if (fid!=0) fclose(fid);
ML_Exit(Comm.MyPID(),"error reading matrix file comments", EXIT_FAILURE);
}
} while (line[0] == '%');
// Next get problem dimensions: M, N, NZ
if(sscanf(line, "%d %d %d", numGlobalRows, &N, &NZ)==0) {
if (fid!=0) fclose(fid);
ML_Exit(Comm.MyPID(),"error reading matrix file dimensions", EXIT_FAILURE);
}
} //ML_Read_Matrix_Dimensions()
void show_matrix(const char *txt, const Epetra_RowMatrix &matrix, const Epetra_Comm &comm)
{
int me = comm.MyPID();
if (comm.NumProc() > 10){
if (me == 0){
std::cout << txt << std::endl;
std::cout << "Printed matrix format only works for 10 or fewer processes" << std::endl;
}
return;
}
int numRows = matrix.NumGlobalRows();
int numCols = matrix.NumGlobalCols();
if ((numRows > 200) || (numCols > 500)){
if (me == 0){
std::cerr << txt << std::endl;
std::cerr << "show_matrix: problem is too large to display" << std::endl;
}
return;
}
int *myA = new int [numRows * numCols];
make_my_A(matrix, myA, comm);
printMatrix(txt, myA, NULL, NULL, numRows, numCols, comm);
delete [] myA;
}
int generateHyprePrintOut(const char *filename, const Epetra_Comm &comm){
int MyPID = comm.MyPID();
int NumProc = comm.NumProc();
int N = 100;
int ilower = MyPID * N;
int iupper = (MyPID+1)*N-1;
double filePID = (double)MyPID/(double)100000;
std::ostringstream stream;
// Using setprecision() puts it in the std::string
stream << std::setiosflags(std::ios::fixed) << std::setprecision(5) << filePID;
// Then just ignore the first character
std::string fileName(filename);
fileName += stream.str().substr(1,7);
std::ofstream myfile(fileName.c_str());
if(myfile.is_open()){
myfile << ilower << " " << iupper << " " << ilower << " " << iupper << std::endl;
for(int i = ilower; i <= iupper; i++){
for(int j=i-5; j <= i+5; j++){
if(j >= 0 && j < N*NumProc)
myfile << i << " " << j << " " << (double)rand()/(double)RAND_MAX << std::endl;
}
}
myfile.close();
return 0;
} else {
std::cout << "\nERROR:\nCouldn't open file.\n";
return -1;
}
}
int special_submap_import_test(Epetra_Comm& Comm)
{
int localProc = Comm.MyPID();
//set up ids_source and ids_target such that ids_source are only
//a subset of ids_target, and furthermore that ids_target are ordered
//such that the LIDs don't match up. In other words, even if gid 2 does
//exist in both ids_source and ids_target, it will correspond to different
//LIDs on at least 1 proc.
//
//This is to test a certain bug-fix in Epetra_Import where the 'RemoteLIDs'
//array wasn't being calculated correctly on all procs.
long long ids_source[1];
ids_source[0] = localProc*2+2;
long long ids_target[3];
ids_target[0] = localProc*2+2;
ids_target[1] = localProc*2+1;
ids_target[2] = localProc*2+0;
Epetra_Map map_source((long long) -1, 1, &ids_source[0], 0LL, Comm);
Epetra_Map map_target((long long) -1, 3, &ids_target[0], 0LL, Comm);
Epetra_Import importer(map_target, map_source);
Epetra_LongLongVector vec_source(map_source);
Epetra_LongLongVector vec_target(map_target);
vec_target.PutValue(0);
//set vec_source's contents so that entry[i] == GID[i].
long long* GIDs = map_source.MyGlobalElements64();
for(int i=0; i<map_source.NumMyElements(); ++i) {
vec_source[i] = GIDs[i];
}
//Import vec_source into vec_target. This should result in the contents
//of vec_target remaining 0 for the entries that don't exist in vec_source,
//and other entries should be equal to the corresponding GID in the map.
vec_target.Import(vec_source, importer, Insert);
GIDs = map_target.MyGlobalElements64();
int test_failed = 0;
//the test passes if the i-th entry in vec_target equals either 0 or
//GIDs[i].
for(int i=0; i<vec_target.MyLength(); ++i) {
if (vec_target[i] != GIDs[i] && vec_target[i] != 0) test_failed = 1;
}
int global_result;
Comm.MaxAll(&test_failed, &global_result, 1);
//If test didn't fail on any procs, global_result should be 0.
//If test failed on any proc, global_result should be 1.
return global_result;
}
void show_matrix(const char *txt, const Epetra_LinearProblem &problem, const Epetra_Comm &comm)
{
int me = comm.MyPID();
if (comm.NumProc() > 10){
if (me == 0){
std::cout << txt << std::endl;
std::cout << "Printed matrix format only works for 10 or fewer processes" << std::endl;
}
return;
}
Epetra_RowMatrix *matrix = problem.GetMatrix();
Epetra_MultiVector *lhs = problem.GetLHS();
Epetra_MultiVector *rhs = problem.GetRHS();
int numRows = matrix->NumGlobalRows();
int numCols = matrix->NumGlobalCols();
if ((numRows > 200) || (numCols > 500)){
if (me == 0){
std::cerr << txt << std::endl;
std::cerr << "show_matrix: problem is too large to display" << std::endl;
}
return;
}
int *myA = new int [numRows * numCols];
make_my_A(*matrix, myA, comm);
int *myX = new int [numCols];
int *myB = new int [numRows];
memset(myX, 0, sizeof(int) * numCols);
memset(myB, 0, sizeof(int) * numRows);
const Epetra_BlockMap &lhsMap = lhs->Map();
const Epetra_BlockMap &rhsMap = rhs->Map();
int base = lhsMap.IndexBase();
for (int j=0; j < lhsMap.NumMyElements(); j++){
int colGID = lhsMap.GID(j);
myX[colGID - base] = me + 1;
}
for (int i=0; i < rhsMap.NumMyElements(); i++){
int rowGID = rhsMap.GID(i);
myB[rowGID - base] = me + 1;
}
printMatrix(txt, myA, myX, myB, numRows, numCols, comm);
delete [] myA;
delete [] myX;
delete [] myB;
}
//==============================================================================
Poisson2dOperator::Poisson2dOperator(int nx, int ny, const Epetra_Comm & comm)
: nx_(nx),
ny_(ny),
useTranspose_(false),
comm_(comm),
map_(0),
numImports_(0),
importIDs_(0),
importMap_(0),
importer_(0),
importX_(0),
Label_(0) {
Label_ = "2D Poisson Operator";
int numProc = comm.NumProc(); // Get number of processors
int myPID = comm.MyPID(); // My rank
if (2*numProc > ny) { // ny must be >= 2*numProc (to avoid degenerate cases)
ny = 2*numProc;
ny_ = ny;
std::cout << " Increasing ny to " << ny << " to avoid degenerate distribution on " << numProc << " processors." << std::endl;
}
int chunkSize = ny/numProc;
int remainder = ny%numProc;
if (myPID+1 <= remainder) chunkSize++; // add on remainder
myny_ = chunkSize;
map_ = new Epetra_Map(-1LL, ((long long)nx)*chunkSize, 0, comm_);
if (numProc>1) {
// Build import GID list to build import map and importer
if (myPID>0) numImports_ += nx;
if (myPID+1<numProc) numImports_ += nx;
if (numImports_>0) importIDs_ = new long long[numImports_];
long long * ptr = importIDs_;
long long minGID = map_->MinMyGID64();
long long maxGID = map_->MaxMyGID64();
if (myPID>0) for (int i=0; i< nx; i++) *ptr++ = minGID - nx + i;
if (myPID+1<numProc) for (int i=0; i< nx; i++) *ptr++ = maxGID + i +1;
// At the end of the above step importIDs_ will have a list of global IDs that are needed
// to compute the matrix multiplication operation on this processor. Now build import map
// and importer
importMap_ = new Epetra_Map(-1LL, numImports_, importIDs_, 0LL, comm_);
importer_ = new Epetra_Import(*importMap_, *map_);
}
}
int alternate_import_constructor_test(Epetra_Comm& Comm) {
int rv=0;
int nodes_per_proc=10;
int numprocs = Comm.NumProc();
int mypid = Comm.MyPID();
// Only run if we have multiple procs & MPI
if(numprocs==0) return 0;
#ifndef HAVE_MPI
return 0;
#endif
// Build Map 1 - linear
Epetra_Map Map1((long long)-1,nodes_per_proc,(long long)0,Comm);
// Build Map 2 - mod striped
std::vector<long long> MyGIDs(nodes_per_proc);
for(int i=0; i<nodes_per_proc; i++)
MyGIDs[i] = (mypid*nodes_per_proc + i) % numprocs;
Epetra_Map Map2((long long)-1,nodes_per_proc,&MyGIDs[0],(long long)0,Comm);
// For testing
Epetra_LongLongVector Source(Map1), Target(Map2);
// Build Import 1 - normal
Epetra_Import Import1(Map2,Map1);
rv = rv|| test_import_gid("Alt test: 2 map constructor",Source,Target, Import1);
// Build Import 2 - no-comm constructor
int Nremote=Import1.NumRemoteIDs();
const int * RemoteLIDs = Import1.RemoteLIDs();
std::vector<int> RemotePIDs(Nremote+1); // I hate you, stl vector....
std::vector<int> AllPIDs;
Epetra_Util::GetPids(Import1,AllPIDs,true);
for(int i=0; i<Nremote; i++) {
RemotePIDs[i]=AllPIDs[RemoteLIDs[i]];
}
Epetra_Import Import2(Import1.TargetMap(),Import1.SourceMap(),Nremote,&RemotePIDs[0],Import1.NumExportIDs(),Import1.ExportLIDs(),Import1.ExportPIDs());
rv = rv || test_import_gid("Alt test: no comm constructor",Source,Target,Import2);
// Build Import 3 - Remotes only
Epetra_Import Import3(Import1.TargetMap(),Import1.SourceMap(),Nremote,&RemotePIDs[0]);
rv = rv || test_import_gid("Alt test: remote only constructor",Source,Target, Import3);
return rv;
}
int combine_mode_test(Epetra_Comm& Comm)
{
int localProc = Comm.MyPID();
long long ids_source[1];
ids_source[0] = localProc*2+2;
long long ids_target[3];
ids_target[0] = localProc*2+2;
ids_target[1] = localProc*2+1;
ids_target[2] = localProc*2+0;
Epetra_Map map_source((long long) -1, 1, &ids_source[0], 0LL, Comm);
Epetra_Map map_target((long long) -1, 3, &ids_target[0], 0LL, Comm);
Epetra_Import importer(map_target, map_source);
Epetra_LongLongVector vec_source(map_source);
Epetra_LongLongVector vec_target(map_target);
vec_target.PutValue(0);
//set vec_source's contents so that entry[i] == GID[i].
long long* GIDs = map_source.MyGlobalElements64();
for(int i=0; i<map_source.NumMyElements(); ++i) {
vec_source[i] = GIDs[i];
}
//Import vec_source into vec_target. This should result in the contents
//of vec_target remaining 0 for the entries that don't exist in vec_source,
//and other entries should be equal to the corresponding GID in the map.
vec_target.Import(vec_source, importer, Insert);
GIDs = map_target.MyGlobalElements64();
int test_failed = 0;
//the test passes if the i-th entry in vec_target equals either 0 or
//GIDs[i].
for(int i=0; i<vec_target.MyLength(); ++i) {
if (vec_target[i] != GIDs[i] && vec_target[i] != 0) test_failed = 1;
}
int global_result;
Comm.MaxAll(&test_failed, &global_result, 1);
//If test didn't fail on any procs, global_result should be 0.
//If test failed on any proc, global_result should be 1.
return global_result;
}
//==============================================================================
// Epetra_BlockMap constructor function for a Epetra-defined uniform linear distribution of constant size elements.
void Epetra_BlockMap::ConstructAutoUniform(long long NumGlobal_Elements, int Element_Size, int Index_Base, const Epetra_Comm& comm, bool IsLongLong)
{
// Each processor gets roughly numGlobalPoints/p points
// This routine automatically defines a linear partitioning of a
// map with numGlobalPoints across the processors
// specified in the given Epetra_Comm
if (NumGlobal_Elements < 0)
throw ReportError("NumGlobal_Elements = " + toString(NumGlobal_Elements) + ". Should be >= 0.", -1);
if (Element_Size <= 0)
throw ReportError("ElementSize = " + toString(Element_Size) + ". Should be > 0.", -2);
BlockMapData_ = new Epetra_BlockMapData(NumGlobal_Elements, Element_Size, Index_Base, comm, IsLongLong);
int NumProc = comm.NumProc();
BlockMapData_->ConstantElementSize_ = true;
BlockMapData_->LinearMap_ = true;
int MyPID = comm.MyPID();
if(BlockMapData_->NumGlobalElements_ / NumProc > (long long) std::numeric_limits<int>::max())
throw ReportError("Epetra_BlockMap::ConstructAutoUniform: Error. Not enough space for elements on each processor", -99);
BlockMapData_->NumMyElements_ = (int) (BlockMapData_->NumGlobalElements_ / NumProc);
int remainder = (int) (BlockMapData_->NumGlobalElements_ % NumProc); // remainder will fit int
int start_index = MyPID * (BlockMapData_->NumMyElements_ + 1);
if (MyPID < remainder)
BlockMapData_->NumMyElements_++;
else
start_index -= (MyPID - remainder);
BlockMapData_->NumGlobalPoints_ = BlockMapData_->NumGlobalElements_ * BlockMapData_->ElementSize_;
BlockMapData_->NumMyPoints_ = BlockMapData_->NumMyElements_ * BlockMapData_->ElementSize_;
BlockMapData_->MinMyElementSize_ = BlockMapData_->ElementSize_;
BlockMapData_->MaxMyElementSize_ = BlockMapData_->ElementSize_;
BlockMapData_->MinElementSize_ = BlockMapData_->ElementSize_;
BlockMapData_->MaxElementSize_ = BlockMapData_->ElementSize_;
BlockMapData_->MinAllGID_ = BlockMapData_->IndexBase_;
BlockMapData_->MaxAllGID_ = BlockMapData_->MinAllGID_ + BlockMapData_->NumGlobalElements_ - 1;
BlockMapData_->MinMyGID_ = start_index + BlockMapData_->IndexBase_;
BlockMapData_->MaxMyGID_ = BlockMapData_->MinMyGID_ + BlockMapData_->NumMyElements_ - 1;
BlockMapData_->DistributedGlobal_ = IsDistributedGlobal(BlockMapData_->NumGlobalElements_, BlockMapData_->NumMyElements_);
EndOfConstructorOps();
}
int rebalanceEpetraProblem( RCP<Epetra_Map> &Map,
RCP<Epetra_CrsMatrix> &A,
RCP<Epetra_MultiVector> &B,
RCP<Epetra_MultiVector> &X,
Epetra_Comm &Comm
)
{
// Rebalance linear system across multiple processors.
if ( Comm.NumProc() > 1 ) {
RCP<Epetra_Map> newMap = rcp( new Epetra_Map( Map->NumGlobalElements(), Map->IndexBase(), Comm ) );
RCP<Epetra_Import> newImport = rcp( new Epetra_Import( *newMap, *Map ) );
// Create rebalanced versions of the linear system.
RCP<Epetra_CrsMatrix> newA = rcp( new Epetra_CrsMatrix( BELOSEPETRACOPY, *newMap, 0 ) );
newA->Import( *A, *newImport, Insert );
newA->FillComplete();
RCP<Epetra_MultiVector> newB = rcp( new Epetra_MultiVector( *newMap, B->NumVectors() ) );
newB->Import( *B, *newImport, Insert );
RCP<Epetra_MultiVector> newX = rcp( new Epetra_MultiVector( *newMap, X->NumVectors() ) );
newX->Import( *X, *newImport, Insert );
// Set the pointers to the new rebalance linear system.
A = newA;
B = newB;
X = newX;
Map = newMap;
}
return (0);
}
Teuchos::RCP<Epetra_CrsMatrix> buildMatrix(int nx, Epetra_Comm & comm)
{
Epetra_Map map(nx*comm.NumProc(),0,comm);
Teuchos::RCP<Epetra_CrsMatrix> mat = Teuchos::rcp(new Epetra_CrsMatrix(Copy,map,3));
int offsets[3] = {-1, 0, 1 };
double values[3] = { -1, 2, -1};
int maxGid = map.MaxAllGID();
for(int lid=0;lid<nx;lid++) {
int gid = mat->GRID(lid);
int numEntries = 3, offset = 0;
int indices[3] = { gid+offsets[0],
gid+offsets[1],
gid+offsets[2] };
if(gid==0) { // left end point
numEntries = 2;
offset = 1;
} // right end point
else if(gid==maxGid)
numEntries = 2;
// insert rows
mat->InsertGlobalValues(gid,numEntries,values+offset,indices+offset);
}
mat->FillComplete();
return mat;
}
int fevec4(Epetra_Comm& Comm, bool verbose)
{
int NumElements = 4;
Epetra_Map Map(NumElements, 0, Comm);
Epetra_FEVector x1(Map);
const double value = 1.;
x1.PutScalar (value);
// replace one element by itself. processor 0
// does not own this element
const int GID = 3;
x1.ReplaceGlobalValues(1, &GID, &value);
x1.GlobalAssemble (Insert);
if (Map.MyGID(3)) {
//insist that the value for GID==3 is 1:
if (std::abs(x1.Values()[Map.LID(3)] - 1) > 1.e-9) return -1;
}
std::cout << x1;
Comm.Barrier();
// re-apply GlobalAssemble. Nothing should
// happen
x1.GlobalAssemble (Insert);
std::cout << x1;
if (Map.MyGID(3)) {
//insist that the value for GID==3 is 1:
if (std::abs(x1.Values()[Map.LID(3)] - 1) > 1.e-9) return -1;
}
return 0;
}
Teuchos::RCP<const EpetraExt::MultiComm>
Stokhos::buildMultiComm(const Epetra_Comm& globalComm,
int num_global_stochastic_blocks,
int num_spatial_procs)
{
Teuchos::RCP<const EpetraExt::MultiComm> globalMultiComm;
#ifdef HAVE_MPI
if (num_spatial_procs == -1) {
// By default, use all procs for spatial parallelism
//MPI_Comm_size(MPI_COMM_WORLD, &num_spatial_procs);
num_spatial_procs = globalComm.NumProc();
}
const Epetra_MpiComm& globalMpiComm =
dynamic_cast<const Epetra_MpiComm&>(globalComm);
globalMultiComm =
Teuchos::rcp(new EpetraExt::MultiMpiComm(globalMpiComm.Comm(),
num_spatial_procs,
num_global_stochastic_blocks,
Teuchos::VERB_NONE));
#else
globalMultiComm =
Teuchos::rcp(new EpetraExt::MultiSerialComm(num_global_stochastic_blocks));
#endif
return globalMultiComm;
}
int HypreFileToCrsMatrix(const char *filename, const Epetra_Comm &comm, Epetra_CrsMatrix *&Matrix){
int MyPID = comm.MyPID();
// This double will be in the format we want for the extension besides the leading zero
double filePID = (double)MyPID/(double)100000;
std::ostringstream stream;
// Using setprecision() puts it in the string
stream << std::setiosflags(std::ios::fixed) << std::setprecision(5) << filePID;
// Then just ignore the first character
std::string fileName(filename);
fileName += stream.str().substr(1,7);
// Open the file
std::ifstream file(fileName.c_str());
string line;
if(file.is_open()){
std::getline(file, line);
int ilower, iupper;
std::istringstream istream(line);
// The first line of the file has the beginning and ending rows
istream >> ilower;
istream >> iupper;
// Using those we can create a row map
Epetra_Map RowMap(-1, iupper-ilower+1, 0, comm);
Matrix = new Epetra_CrsMatrix(Copy, RowMap, 0);
int currRow = -1;
int counter = 0;
std::vector<int> indices;
std::vector<double> values;
while(!file.eof()){
std::getline(file, line);
std::istringstream lineStr(line);
int row, col;
double val;
lineStr >> row;
lineStr >> col;
lineStr >> val;
if(currRow == -1) currRow = row; // First line
if(row == currRow){
// add to the vector
counter = counter + 1;
indices.push_back(col);
values.push_back(val);
} else {
Matrix->InsertGlobalValues(currRow, counter, &values[0], &indices[0]);
indices.clear();
values.clear();
counter = 0;
currRow = row;
// make a new vector
indices.push_back(col);
values.push_back(val);
counter = counter + 1;
}
}
Matrix->InsertGlobalValues(currRow, counter, &values[0], &indices[0]);
Matrix->Comm().Barrier();
Matrix->FillComplete();
file.close();
return 0;
} else {
Teuchos::RCP< Epetra_LinearProblem >
build_problem(Teuchos::ParameterList& test_params,
const Epetra_Comm& comm)
{
Teuchos::Time timer("build_problem");
timer.start();
Epetra_CrsMatrix* A;
Epetra_Vector* b = NULL;
std::string mm_file("not specified");
std::string rhs_mm_file("not specified");
helper::GetParameter(test_params, "mm_file", mm_file);
helper::GetParameter(test_params, "rhs_mm_file", rhs_mm_file);
std::string hb_file("not specified");
helper::GetParameter(test_params, "hb_file", hb_file);
if (mm_file != "not specified") {
if (comm.MyPID() == 0) {
std::cout << "Matrix-Market file: " << mm_file << std::endl;
}
A = read_matrix_mm(mm_file, comm);
if (rhs_mm_file != "not specified") {
if (comm.MyPID() == 0) {
std::cout << "Matrix-Market file: " << rhs_mm_file << std::endl;
}
b = read_vector_mm(rhs_mm_file, comm);
}
}
else if (hb_file != "not specified") {
read_matrix_hb(hb_file, comm, A, b);
}
else {
throw std::runtime_error("No matrix file specified.");
}
Teuchos::RCP<Epetra_LinearProblem> problem = build_problem_mm(test_params, A, b);
timer.stop();
if (comm.MyPID() == 0) {
std::cout << "proc 0 time to read matrix & create problem: " << timer.totalElapsedTime()
<< std::endl;
}
return problem;
}
static void print_out(const Epetra_Comm& Comm, const int level, const char* what)
{
if (Comm.MyPID() == 0 && ML_Get_PrintLevel() > 2)
#ifdef TFLOP
printf("Amesos (level %d) : Building %s\n", level, what);
#else
std::cout << "Amesos (level " << level << ") : Building " << what << "\n";
#endif
}
// Do something with the given communicator. In this case, we just
// print Epetra's version to the given output stream, on Process 0.
void
exampleRoutine (const Epetra_Comm& comm,
std::ostream& out)
{
if (comm.MyPID () == 0) {
// On (MPI) Process 0, print out the Epetra software version.
out << Epetra_Version () << std::endl << std::endl;
}
}
void build_maps(
itype nrows, // Number of global matrix rows
bool testEpetra64,// Flag indicating whether to adjust global row/column
// indices to exercise Epetra64 capability.
Epetra_Comm &comm, // Epetra communicator to be used in maps
Epetra_Map **vectorMap, // OUTPUT: Map to be used for the vector
Epetra_Map **rowMap, // OUTPUT: Map to be used for the matrix rows
Epetra_Map **colMap, // OUTPUT: Map to be used for the matrix cols
long long &offsetEpetra64, // OUTPUT for testing Epetra64: add offsetEpetra64
// to all row/column indices.
bool verbose // print out generated maps
)
{
// Function to build the maps for 1D or 2D matrix distribution.
// Output for 1D includes rowMap and NULL colMap and vectorMap.
// Output for 2D includes rowMap, colMap and vectorMap.
int me = comm.MyPID();
int np = comm.NumProc();
*rowMap = NULL;
*colMap = NULL;
*vectorMap = NULL;
// offsetEpetra64 = (testEpetra64 ? (long long) INT_MAX - (long long) 5 : 0);
offsetEpetra64 = (testEpetra64 ? (long long) 2 * INT_MAX : 0);
// Generate 1D row-based decomposition.
if ((me == 0) && verbose)
cout << endl
<< "1D Distribution: " << endl
<< " np = " << np << endl;
// Linear map similar to Trilinos default.
itype nMyRows = nrows / np + (nrows % np > me);
itype myFirstRow = me * (nrows / np) + MIN(nrows % np, me);
itype *myGlobalRows = new itype[nMyRows];
for (itype i = 0; i < nMyRows; i++)
myGlobalRows[i] = i + myFirstRow + offsetEpetra64;
*rowMap = new Epetra_Map(nrows, nMyRows, &myGlobalRows[0], 0, comm);
delete [] myGlobalRows;
}
int rectangular(const Epetra_Comm& Comm, bool verbose)
{
int mypid = Comm.MyPID();
int numlocalrows = 3;
Epetra_Map rowmap((long long) -1, numlocalrows, 0, Comm);
long long numglobalrows = numlocalrows*Comm.NumProc();
long long numcols = 2*numglobalrows;
Epetra_FECrsGraph fegraph(Copy, rowmap, numcols);
long long* cols = new long long[numcols];
for(int j=0; j<numcols; ++j) cols[j] = j;
Epetra_Map domainmap((long long) -1, numcols, 0, Comm);
long long firstlocalrow = numlocalrows*mypid;
long long lastlocalrow = numlocalrows*(mypid+1)-1;
for(long long i=0; i<numglobalrows; ++i) {
//if i is a local row, then skip it. We want each processor to only
//load rows that belong on other processors.
if (i >= firstlocalrow && i <= lastlocalrow) continue;
EPETRA_CHK_ERR( fegraph.InsertGlobalIndices(1, &i, numcols, &(cols[0])) );
}
EPETRA_CHK_ERR( fegraph.GlobalAssemble(domainmap, rowmap) );
if (verbose) {
std::cout << "********************** fegraph **********************" << std::endl;
std::cout << fegraph << std::endl;
}
delete [] cols;
return(0);
}
void MPIWrapper::allGatherCompact(const Epetra_Comm &Comm, FieldContainer<Scalar> &gatheredValues,
FieldContainer<Scalar> &myValues, FieldContainer<int> &offsets)
{
int mySize = myValues.size();
int totalSize;
Comm.SumAll(&mySize, &totalSize, 1);
int myOffset = 0;
Comm.ScanSum(&mySize,&myOffset,1);
myOffset -= mySize;
gatheredValues.resize(totalSize);
for (int i=0; i<mySize; i++)
{
gatheredValues[myOffset+i] = myValues[i];
}
MPIWrapper::entryWiseSum(Comm, gatheredValues);
offsets.resize(Comm.NumProc());
offsets[Comm.MyPID()] = myOffset;
MPIWrapper::entryWiseSum(Comm, offsets);
}
void MPIWrapper::allGatherHomogeneous(const Epetra_Comm &Comm, FieldContainer<int> &allValues, FieldContainer<int> &myValues)
{
int numProcs = Teuchos::GlobalMPISession::getNProc();
if (numProcs != allValues.dimension(0))
{
TEUCHOS_TEST_FOR_EXCEPTION(true, std::invalid_argument, "allValues first dimension must be #procs");
}
if (allValues.size() / numProcs != myValues.size())
{
TEUCHOS_TEST_FOR_EXCEPTION(true, std::invalid_argument, "myValues size invalid");
}
#ifdef HAVE_MPI
Comm.GatherAll(&myValues[0], &allValues[0], allValues.size()/numProcs);
#else
#endif
}
static int make_my_A(const Epetra_RowMatrix &matrix, int *myA, const Epetra_Comm &comm)
{
int me = comm.MyPID();
const Epetra_Map &rowmap = matrix.RowMatrixRowMap();
const Epetra_Map &colmap = matrix.RowMatrixColMap();
int myRows = matrix.NumMyRows();
int numRows = matrix.NumGlobalRows();
int numCols = matrix.NumGlobalCols();
int base = rowmap.IndexBase();
int maxRow = matrix.MaxNumEntries();
memset(myA, 0, sizeof(int) * numRows * numCols);
int *myIndices = new int [maxRow];
double *tmp = new double [maxRow];
int rowLen = 0;
for (int i=0; i< myRows; i++){
int rc = matrix.ExtractMyRowCopy(i, maxRow, rowLen, tmp, myIndices);
if (rc){
if (me == 0){
std::cout << "Error in make_my_A" << std::endl;
}
return 1;
}
int *row = myA + (numCols * (rowmap.GID(i) - base));
for (int j=0; j < rowLen; j++){
int colGID = colmap.GID(myIndices[j]);
row[colGID - base] = me + 1;
}
}
if (maxRow){
delete [] myIndices;
delete [] tmp;
}
return 0;
}
int fevec7(Epetra_Comm& Comm, bool verbose)
{
const int NumVectors = 4;
const int NumElements = 4;
Epetra_Map Map(NumElements, 0, Comm);
std::vector<double> mydata(NumElements*NumVectors, 1.0);
Epetra_FEVector x1(View, Map, &mydata[0], NumElements, NumVectors);
x1.PutScalar (0);
// let all processors set global entry 0 to 1
const int GID = 0;
const double value = 1;
x1.ReplaceGlobalValues(1, &GID, &value);
x1.GlobalAssemble (Insert);
if (Comm.MyPID()==0 && x1[0][0] != value) return -1;
return 0;
}
int check_graph_sharing(Epetra_Comm& Comm)
{
int numLocalElems = 5;
int localProc = Comm.MyPID();
int firstElem = localProc*numLocalElems;
int err;
Epetra_Map map(-1, numLocalElems, 0, Comm);
Epetra_CrsMatrix* A = new Epetra_CrsMatrix(Copy, map, 1);
for (int i=0; i<numLocalElems; ++i) {
int row = firstElem+i;
int col = row;
double val = 1.0;
err = A->InsertGlobalValues(row, 1, &val, &col);
if (err != 0) {
cerr << "A->InsertGlobalValues("<<row<<") returned err="<<err<<endl;
return(err);
}
}
A->FillComplete(false);
Epetra_CrsMatrix B(Copy, A->Graph());
delete A;
for (int i=0; i<numLocalElems; ++i) {
int row = firstElem+i;
int col = row;
double val = 1.0;
err = B.ReplaceGlobalValues(row, 1, &val, &col);
if (err != 0) {
cerr << "B.InsertGlobalValues("<<row<<") returned err="<<err<<endl;
return(err);
}
}
return(0);
}
int test_bug2890(Epetra_Comm& Comm, bool verbose)
{
//This function tests the AZ_random1() function in AztecOO.
//The implementation of the Park and Miller random number
//generator was incorrect and resulted in an overflow condition.
//This is *not* a complete test of AztecOO's RNG.
//
//A more robust check is to compile AztecOO with gcc -ftrapv and run
//a Krylov method that invokes AZ_random_vector().
int seed = -127773;
double rand_num;
rand_num = AZ_srandom1(&seed);
if (verbose && Comm.MyPID() == 0)
printf("test_bug2890: rand_num = %e (should be in [0,1])\n",rand_num);
if ( (rand_num > 1) || (rand_num < 0) )
return 1; // rand_num should be in [0,1]
else
return 0;
}
//============================================================================
void Ifpack_BreakForDebugger(Epetra_Comm& Comm)
{
char hostname[80];
char buf[80];
if (Comm.MyPID() == 0) cout << "Host and Process Ids for tasks" << endl;
for (int i = 0; i <Comm.NumProc() ; i++) {
if (i == Comm.MyPID() ) {
#if defined(TFLOP) || defined(JANUS_STLPORT)
sprintf(buf, "Host: %s PID: %d", "janus", getpid());
#elif defined(_WIN32)
sprintf(buf,"Windows compiler, unknown hostname and PID!");
#else
gethostname(hostname, sizeof(hostname));
sprintf(buf, "Host: %s\tComm.MyPID(): %d\tPID: %d",
hostname, Comm.MyPID(), getpid());
#endif
printf("%s\n",buf);
fflush(stdout);
#if !( defined(_WIN32) )
sleep(1);
#endif
}
}
if(Comm.MyPID() == 0) {
printf("\n");
printf("** Pausing to attach debugger...\n");
printf("** You may now attach debugger to the processes listed above.\n");
printf( "**\n");
printf( "** Enter a character to continue > "); fflush(stdout);
char go;
scanf("%c",&go);
}
Comm.Barrier();
}
void Trilinos_Util_distrib_vbr_matrix(const Epetra_Comm & Comm,
int *N_global, int *N_blk_global,
int *n_nonzeros, int *n_blk_nonzeros,
int *N_update, int **update,
double **val, int **indx, int **rpntr, int **cpntr,
int **bpntr, int **bindx,
double **x, double **b, double **xexact)
#undef DEBUG
{
int i, n_global_nonzeros, n_global_blk_nonzeros;
int N_local;
int j, row, have_xexact = 0 ;
int *rpntr1, *bindx1, *bpntr1, *indx1;
double *val1, *b1, *x1, *xexact1=0;
int MyPID = Comm.MyPID();
int NumProc = Comm.NumProc();
printf("Processor %d of %d entering distrib_matrix.\n",
MyPID,NumProc) ;
/*************** Distribute global matrix to all processors ************/
if(MyPID == 0)
{
if ((*xexact) != NULL) have_xexact = 1;
printf("%s", "Broadcasting exact solution\n");
}
if(NumProc > 1)
{
Comm.Broadcast( N_global, 1, 0);
Comm.Broadcast( N_blk_global, 1, 0);
Comm.Broadcast( n_nonzeros, 1, 0);
Comm.Broadcast( n_blk_nonzeros, 1, 0);
Comm.Broadcast( &have_xexact, 1, 0);
printf("Processor %d of %d done with global parameter broadcast.\n",
MyPID,NumProc) ;
if(MyPID != 0)
{
*bpntr = (int *) calloc(*N_blk_global+1,sizeof(int)) ;
*rpntr = (int *) calloc(*N_blk_global+1,sizeof(int)) ;
*bindx = (int *) calloc(*n_blk_nonzeros+1,sizeof(int)) ;
*indx = (int *) calloc(*n_blk_nonzeros+1,sizeof(int)) ;
*val = (double *) calloc(*n_nonzeros+1,sizeof(double)) ;
printf("Processor %d of %d done with global calloc.\n",
MyPID,NumProc) ;
}
Comm.Broadcast( (*bpntr), (*N_blk_global+1), 0);
Comm.Broadcast( (*rpntr), (*N_blk_global+1), 0);
Comm.Broadcast( (*bindx), (*n_blk_nonzeros+1), 0);
Comm.Broadcast( (*indx), (*n_blk_nonzeros+1), 0);
Comm.Broadcast( (*val), (*n_nonzeros+1), 0);
printf("Processor %d of %d done with matrix broadcast.\n",
MyPID,NumProc) ;
/* Set rhs and initialize guess */
if(MyPID != 0)
{
(*b) = (double *) calloc(*N_global,sizeof(double)) ;
(*x) = (double *) calloc(*N_global,sizeof(double)) ;
if (have_xexact)
(*xexact) = (double *) calloc(*N_global,sizeof(double)) ;
}
Comm.Broadcast( (*x), (*N_global), 0);
Comm.Broadcast( (*b), (*N_global), 0);
if (have_xexact)
Comm.Broadcast((*xexact), (*N_global), 0);
printf("Processor %d of %d done with rhs/guess broadcast.\n",
MyPID,NumProc) ;
}
/********************** Generate update map *************************/
//read_update(N_update, update, proc_config, *N_blk_global, 1, linear) ;
Epetra_Map map(*N_blk_global, 0, Comm);
*N_update = map.NumMyElements();
(*update) = (int *) calloc(*N_update,sizeof(int)) ;
map.MyGlobalElements(*update);
printf("Processor %d of %d has %d rows of %d total block rows.\n",
MyPID,NumProc,*N_update,*N_blk_global) ;
/*************** Construct local matrix from global matrix ************/
/* The local matrix is a copy of the rows assigned to this processor.
It is stored in MSR format and still has global indices
*/
if(NumProc > 1)
{
n_global_nonzeros = *n_nonzeros;
n_global_blk_nonzeros = *n_blk_nonzeros;
*n_nonzeros = 0;
*n_blk_nonzeros = 0;
N_local = 0;
for (i=0; i<*N_update; i++)
{
row = (*update)[i];
*n_nonzeros += (*indx)[(*bpntr)[row+1]] - (*indx)[(*bpntr)[row]];
*n_blk_nonzeros += (*bpntr)[row+1] - (*bpntr)[row];
N_local += (*rpntr)[row+1] - (*rpntr)[row];
}
printf("Processor %d of %d has %d nonzeros of %d total nonzeros.\n",
MyPID,NumProc,
*n_nonzeros,n_global_nonzeros) ;
printf("Processor %d of %d has %d block nonzeros of %d total block nonzeros.\n",
MyPID,NumProc,
*n_blk_nonzeros,n_global_blk_nonzeros) ;
printf("Processor %d of %d has %d equations of %d total equations.\n",
MyPID,NumProc,
N_local,*N_global) ;
#ifdef DEBUG
{ double sum1 = 0.0;
for (i=0;i<*N_global; i++) sum1 += (*b)[i];
printf("Processor %d of %d has sum of b = %12.4g.\n",
MyPID,NumProc,sum1) ;
}
#endif /* DEBUG */
/* Allocate memory for local matrix */
bpntr1 = (int *) calloc(*N_update+1,sizeof(int)) ;
rpntr1 = (int *) calloc(*N_update+1,sizeof(int)) ;
bindx1 = (int *) calloc(*n_blk_nonzeros+1,sizeof(int)) ;
indx1 = (int *) calloc(*n_blk_nonzeros+1,sizeof(int)) ;
val1 = (double *) calloc(*n_nonzeros+1,sizeof(double)) ;
b1 = (double *) calloc(N_local,sizeof(double)) ;
x1 = (double *) calloc(N_local,sizeof(double)) ;
if (have_xexact)
xexact1 = (double *) calloc(N_local,sizeof(double)) ;
{
int cur_blk_size, indx_offset, len_val, row_offset, row_offset1;
double *val_ptr, *val1_ptr;
bpntr1[0] = 0;
indx1[0] = 0;
rpntr1[0] = 0;
for (i=0; i<*N_update; i++)
{
row = (*update)[i];
cur_blk_size = (*rpntr)[row+1] - (*rpntr)[row];
rpntr1[i+1] = rpntr1[i] + cur_blk_size;
row_offset = (*rpntr)[row];
row_offset1 = rpntr1[i];
for (j = 0; j<cur_blk_size; j++)
{
b1[row_offset1+j] = (*b)[row_offset+j];
x1[row_offset1+j] = (*x)[row_offset+j];
if (have_xexact) xexact1[row_offset1+j] = (*xexact)[row_offset+j];
}
bpntr1[i+1] = bpntr1[i];
#ifdef DEBUG
printf("Proc %d of %d: Global row = %d: Local row = %d: b = %12.4g: x = %12.4g: bindx = %d: val = %12.4g \n",
MyPID,NumProc,
row, i, b1[i], x1[i], bindx1[i], val1[i]) ;
#endif
indx_offset = (*indx)[(*bpntr)[row]] - indx1[bpntr1[i]];
for (j = (*bpntr)[row]; j < (*bpntr)[row+1]; j++)
{
indx1[bpntr1 [i+1] + 1] = (*indx)[j+1] - indx_offset;
bindx1[bpntr1 [i+1] ] = (*bindx)[j];
bpntr1[i+1] ++;
}
len_val = indx1[bpntr1[i+1]] - indx1[bpntr1[i]];
val_ptr = (*val)+(*indx)[(*bpntr)[row]];
val1_ptr = val1+indx1[bpntr1[i]];
for (j = 0; j<len_val; j++)
{
*val1_ptr = *val_ptr;
val_ptr++; val1_ptr++;
}
}
}
printf("Processor %d of %d done with extracting local operators.\n",
MyPID,NumProc) ;
if (have_xexact)
{
printf(
"The residual using VBR format and exact solution on processor %d is %12.4g\n",
MyPID,
Trilinos_Util_svbrres (N_local, *N_global, *N_update, val1, indx1, bindx1,
rpntr1, (*rpntr), bpntr1, bpntr1+1,
(*xexact), b1));
}
/* Release memory for global matrix, rhs and solution */
free ((void *) (*val));
free ((void *) (*indx));
free ((void *) (*bindx));
free ((void *) (*bpntr));
free ((void *) (*rpntr));
free ((void *) (*b));
free ((void *) (*x));
if (have_xexact) free((void *) *xexact);
/* Return local matrix through same pointers. */
*val = val1;
*indx = indx1;
*bindx = bindx1;
*bpntr = bpntr1;
*rpntr = rpntr1;
*b = b1;
*x = x1;
if (have_xexact) *xexact = xexact1;
}
if (have_xexact && NumProc == 1)
{
printf(
"The residual using VBR format and exact solution on processor %d is %12.4g\n",
MyPID,
Trilinos_Util_svbrres (*N_global, *N_global, *N_update, (*val), (*indx), (*bindx),
(*rpntr), (*rpntr), (*bpntr), (*bpntr)+1,
(*xexact), (*b)));
}
printf("Processor %d of %d leaving distrib_matrix.\n",
MyPID,NumProc) ;
/* end distrib_matrix */
}