int MAIN__()
{
int argc = 1;
char * name = "c_example";
char ** argv;
#else
int main(int argc, char ** argv)
{
#endif
DMUMPS_STRUC_C id;
MUMPS_INT n = 2;
MUMPS_INT nz = 2;
MUMPS_INT irn[] = { 1, 2 };
MUMPS_INT jcn[] = { 1, 2 };
double a[2];
double rhs[2];
MUMPS_INT myid, ierr;
#if defined(MAIN_COMP)
argv = &name;
#endif
ierr = MPI_Init(&argc, &argv);
ierr = MPI_Comm_rank(MPI_COMM_WORLD, &myid);
/* Define A and rhs */
rhs[0] = 1.0; rhs[1] = 4.0;
a[0] = 1.0; a[1] = 2.0;
/* Initialize a MUMPS instance. Use MPI_COMM_WORLD */
id.job = JOB_INIT; id.par = 1; id.sym = 0; id.comm_fortran = USE_COMM_WORLD;
dmumps_c(&id);
/* Define the problem on the host */
if (myid == 0) {
id.n = n; id.nz = nz; id.irn = irn; id.jcn = jcn;
id.a = a; id.rhs = rhs;
}
#define ICNTL(I) icntl[(I)-1] /* macro s.t. indices match documentation */
/* No outputs */
id.ICNTL(1) = -1; id.ICNTL(2) = -1; id.ICNTL(3) = -1; id.ICNTL(4) = 0;
/* Call the MUMPS package. */
id.job = 6;
dmumps_c(&id);
id.job = JOB_END; dmumps_c(&id); /* Terminate instance */
if (myid == 0) {
printf("Solution is : (%8.2f %8.2f)\n", rhs[0], rhs[1]);
}
ierr = MPI_Finalize();
getchar();
return 0;
}
int MAIN__()
{
int argc=1;
char * name = "c_example";
char ** argv ;
#else
int main(int argc, char ** argv)
{
#endif
DMUMPS_STRUC_C id;
int n = 6;
int nz = 21;
int irnL[] = {1,2,3,4,5,6,2,3,4,5,6,3,4,5,6,4,5,6,5,6,6};
int jcnL[] = {1,1,1,1,1,1,2,2,2,2,2,3,3,3,3,4,4,4,5,5,6};
int irnS[] = {1,2,3,2,3,3};
int jcnS[] = {1,1,1,2,2,3};
int *irn = irnL;
int *jcn = jcnL;
double a[21];
double rhs[6];
int myid, ierr, numP;
#if defined(MAIN_COMP)
argv = &name;
#endif
ierr = MPI_Init(&argc, &argv);
ierr = MPI_Comm_rank(MPI_COMM_WORLD, &myid);
ierr = MPI_Comm_size(MPI_COMM_WORLD, &numP);
double nump = numP * 1.0;
/* Define A and rhs */
rhs[0]=2.0;rhs[1]=2.0;rhs[2]=2.0;rhs[3]=2.0;rhs[4]=2.0;rhs[5]=2.0;
a[0]=4169.95/nump;
a[1]=0.0;
a[2]=10075.0/nump;
a[3]=-4030;
a[4]=0.0;
a[5]=0.0;
a[6]=10084.0/nump;
a[7]=-1612.0/nump;
a[8]=0.0;
a[9]=-8.9556;
a[10]=-1612;
a[11]=1354080.0/nump;
a[12]=0.0;
a[13]=1612.0;
a[14]=193440.0;
a[15]=4169.93;
a[16]=0.0;
a[17]=10075;
a[18]=10084;
a[19]=1612;
a[20]=1354000;
if (myid != 0) {
nz = 6;
a[0]=4169.95/nump;
a[1]=0.0;
a[2]=10075.0/nump;
a[3]=10084.0/nump;
a[4]=-1612.0/nump;
a[5]=1354080.0/nump;
irn = irnS;
jcn = jcnS;
}
#define ICNTL(I) icntl[(I)-1] /* macro s.t. indices match documentation */
/* Initialize a MUMPS instance. Use MPI_COMM_WORLD */
id.job=JOB_INIT; id.par=1; id.sym=2; id.comm_fortran=USE_COMM_WORLD;
/* parallel solver; distributed i/p matrix A */
id.ICNTL(5)=0; id.ICNTL(18)=3;
dmumps_c(&id);
/* Define the problem on the host */
/*
id.n = n; id.nz =nz; id.irn=irn; id.jcn=jcn;
id.a = a; id.rhs = rhs;
*/
/* parallel solver; distributed i/p matrix A */
id.ICNTL(5)=0; id.ICNTL(18)=3;
id.n = n; id.nz_loc =nz; id.irn_loc=irn; id.jcn_loc=jcn;
id.a_loc = a; id.rhs = rhs;
/* No outputs */
id.ICNTL(1)=-1; id.ICNTL(2)=-1; id.ICNTL(3)=-1; id.ICNTL(4)=0;
id.job=1;
dmumps_c(&id);
id.job=5;
dmumps_c(&id);
if (myid == 0) {
printf("Solution is : (%8.6e %8.6e %8.6e %8.6e %8.6e %8.6e \n", rhs[0],rhs[1], rhs[2], rhs[3], rhs[4], rhs[5]);
}
rhs[0]=2.0;rhs[1]=2.0;rhs[2]=2.0;rhs[3]=1.0;rhs[4]=1.0;rhs[5]=1.0;
id.job=3;
dmumps_c(&id);
/* Terminate instance */
id.job=JOB_END;
dmumps_c(&id);
if (myid == 0) {
printf("Solution is : (%8.6e %8.6e %8.6e %8.6e %8.6e %8.6e \n", rhs[0],rhs[1], rhs[2], rhs[3], rhs[4], rhs[5]);
}
ierr = MPI_Finalize();
return 0;
}
/* Solve Mx = b */
DATA* yohan::matrix::solveLinearEquation(SquareSparseMatrix* M, DATA* b)
{
//for SymmetricUpperSquareSparseMatrix, using MUMPS
if (M->getType() == SquareSparseMatrix::TYPE_SymmetricMumpsSquareSparse)
{
/* Platform MUMPS relatives*/
DMUMPS_STRUC_C id;
/* JOB_INIT */
id.job = -1; // step: JOB_INIT
id.par = 1; // the host will participate the factorisation/solve phase, in SEQ version, 1 is better
id.sym = 0; // 2 - general symmetric; 1 - symmetric positive definite; 0 - unsymmetric
//id.comm_fortran = -987654;
dmumps_c(&id);
/* Define the problem on the host */
SymmetricMumpsSquareSparseMatrix* A = (SymmetricMumpsSquareSparseMatrix*)M;
id.n = A->order; //order
id.nz = A->size; //count of non-element
id.irn = A->rows;
id.jcn = A->columns;
id.a = A->values;
id.rhs = b;
/* Configuration of the solver */
//id.ICNTL(5) = 0; // the matrix is defined in an "assembled format", (i, j, val)
id.ICNTL(1) = -1;
id.ICNTL(2) = -1;
id.ICNTL(3) = -1;
id.ICNTL(4) = 0; // all these are for the dignostic display, put on -1 if not necessary
/* Call the solver */
id.job = 6;
dmumps_c(&id);
/* We should consider here whether we desallocate the M and the b */
/* JOB_END */
id.job=-2;
dmumps_c(&id);
/* Return */
return b;
}
else if (M->getType() == SquareSparseMatrix::TYPE_SymmetricMumpsSquareSparse2)
{
/* Platform MUMPS relatives*/
DMUMPS_STRUC_C id;
/* JOB_INIT */
id.job = -1; // step: JOB_INIT
id.par = 1; // the host will participate the factorisation/solve phase, in SEQ version, 1 is better
id.sym = 0; // 2 - general symmetric; 1 - symmetric positive definite; 0 - unsymmetric
//id.comm_fortran = -987654;
dmumps_c(&id);
/* Define the problem on the host */
SymmetricMumpsSquareSparseMatrix2* A = (SymmetricMumpsSquareSparseMatrix2*)M;
id.n = A->order; //order
id.nz = A->size; //count of non-element
id.irn = A->rows;
id.jcn = A->columns;
id.a = A->values;
id.rhs = b;
/* Configuration of the solver */
//id.ICNTL(5) = 0; // the matrix is defined in an "assembled format", (i, j, val)
id.ICNTL(1) = -1;
id.ICNTL(2) = -1;
id.ICNTL(3) = -1;
id.ICNTL(4) = 0; // all these are for the dignostic display, put on -1 if not necessary
/* Call the solver */
id.job = 6;
dmumps_c(&id);
/* We should consider here whether we desallocate the M and the b */
/* JOB_END */
id.job=-2;
dmumps_c(&id);
/* Return */
return b;
}
else
return NULL;
//if (A->getType() == SquareSparseMatrix::TYPE_SymmetricIntelSquareSparse)
}
//-----------------------------------------------------------------------------
std::size_t MUMPSLUSolver::solve(GenericVector& x, const GenericVector& b)
{
dolfin_assert(_matA);
DMUMPS_STRUC_C data;
data.comm_fortran = -987654;
// Initialise
data.job = -1;
// Host participates in solve
data.par = 1;
// Output related parameters
//data.ICNTL(1) = 6; // error messages
//data.ICNTL(2) = 0;
//data.ICNTL(3) = 6; // Global information
//data.ICNTL(3) = 6; // Global information
if (parameters["verbose"])
data.ICNTL(4) = 2;
else
data.ICNTL(4) = 1;
// Matrix symmetry (0=non-symmetric, 2=symmetric positive defn, 2=symmetric)
data.sym = 0;
if (parameters["symmetric"])
data.sym = 2;
// Initialise MUMPS
dmumps_c(&data);
// Related to use of ScaLAPACK (+/-. Negative is faster?)
//data.ICNTL(13) = -1;
// Solve transpose (1: A x = b, otherwise A^T x = b)
data.ICNTL(9) = 1;
// FIXME (20=default)
data.ICNTL(14) = 20;
// Reordering (7=automatic)
data.ICNTL(7) = 7;
// Control solution vector (0=solution on root, 1=solution distributed)
data.ICNTL(21) = 1;
// Distributed matrix
data.ICNTL(18) = 3;
// Parallel/serial analysis (0=auto, 1=serial, 2=parallel)
if (MPI::size(_matA->mpi_comm()) > 1)
data.ICNTL(28) = 2;
else
data.ICNTL(28) = 0;
// Parallel graph partitioning library (0=auto, 1=pt-scotch, 2=parmetis)
data.ICNTL(29) = 0;
// Global size
dolfin_assert(_matA->size(0) == _matA->size(1));
data.n = _matA->size(0);
if (!_matA->base_one())
dolfin_error("MUMPSLUSolver.cpp",
"initialize solver",
"MUMPS requires a CoordinateMatrix with Fortran-style "
"base 1 indexing");
// Get matrix coordinate and value data
const std::vector<std::size_t>& rows = _matA->rows();
const std::vector<std::size_t>& cols = _matA->columns();
const std::vector<double>& vals = _matA->values();
// Number of non-zero entries on this process
data.nz_loc = rows.size();
// Pass matrix data to MUMPS. Trust MUMPS not to change it
data.irn_loc = const_cast<int*>(reinterpret_cast<const int*>(rows.data()));
data.jcn_loc = const_cast<int*>(reinterpret_cast<const int*>(cols.data()));
data.a_loc = const_cast<double*>(vals.data());
// Analyse and factorize
data.job = 4;
dmumps_c(&data);
if (data.INFOG(1) < 0)
dolfin_error("MUMPSLUSolver.cpp",
"compute matrix factors",
"MUMPS reported an error during the analysis and "
"factorisation");
cout << "Factorisation finished" << endl;
// Gather RHS on root process and attach
std::vector<double> _b;
b.gather_on_zero(_b);
data.rhs = _b.data();
// Scaling strategy (77 is default)
data.ICNTL(8) = 77;
// Get size of local solution vector x and create objects to hold solution
const std::size_t local_x_size = data.INFO(23);
std::vector<int> x_local_indices(local_x_size);
std::vector<double> x_local_vals(local_x_size);
// Attach solution data to MUMPS object
data.lsol_loc = local_x_size;
data.sol_loc = x_local_vals.data();
data.isol_loc = x_local_indices.data();
// Solve problem
data.job = 3;
dmumps_c(&data);
if (data.INFOG(1) < 0)
dolfin_error("MUMPSLUSolver.cpp",
"compute matrix factors",
"MUMPS reported an error during the solve");
// Shift indices by -1
for (std::size_t i = 0; i < local_x_size ; ++i)
x_local_indices[i]--;
// Set x values
#if defined(PETSC_USE_64BIT_INDICES)
// Cast indices to 64 bit
std::vector<dolfin::la_index> _x_local_indices(x_local_indices.begin(),
x_local_indices.end());
x.set_local(x_local_vals.data(), x_local_indices.size(),
_x_local_indices.data());
#else
x.set_local(x_local_vals.data(), x_local_indices.size(),
x_local_indices.data());
#endif
x.apply("insert");
// Clean up
data.job = -2;
dmumps_c(&data);
return 1;
}
