Homec4science
Diffusion akantu (features/30-implementing-a-checkpoint-restart-mechanism) Browse

rAKA/src/solver/sparse_solver_mumps.cc
082831982c3ffeatures/30-implementin…

Last Change
Raw File

sparse_solver_mumps.cc
View Options

/**
* @file sparse_solver_mumps.cc
*
* @author Nicolas Richart <nicolas.richart@epfl.ch>
*
* @date creation: Mon Dec 13 2010
* @date last modification: Tue Feb 20 2018
*
* @brief implem of SparseSolverMumps class
*
* @section LICENSE
*
* Copyright (©) 2010-2018 EPFL (Ecole Polytechnique Fédérale de Lausanne)
* Laboratory (LSMS - Laboratoire de Simulation en Mécanique des Solides)
*
* Akantu is free software: you can redistribute it and/or modify it under the
* terms of the GNU Lesser General Public License as published by the Free
* Software Foundation, either version 3 of the License, or (at your option) any
* later version.
*
* Akantu is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
* A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License
* along with Akantu. If not, see <http://www.gnu.org/licenses/>.
*
* @section DESCRIPTION
*
* @subsection Ctrl_param Control parameters
*
* ICNTL(1),
* ICNTL(2),
* ICNTL(3) : output streams for error, diagnostics, and global messages
*
* ICNTL(4) : verbose level : 0 no message - 4 all messages
*
* ICNTL(5) : type of matrix, 0 assembled, 1 elementary
*
* ICNTL(6) : control the permutation and scaling(default 7) see mumps doc for
* more information
*
* ICNTL(7) : determine the pivot order (default 7) see mumps doc for more
* information
*
* ICNTL(8) : describe the scaling method used
*
* ICNTL(9) : 1 solve A x = b, 0 solve At x = b
*
* ICNTL(10) : number of iterative refinement when NRHS = 1
*
* ICNTL(11) : > 0 return statistics
*
* ICNTL(12) : only used for SYM = 2, ordering strategy
*
* ICNTL(13) :
*
* ICNTL(14) : percentage of increase of the estimated working space
*
* ICNTL(15-17) : not used
*
* ICNTL(18) : only used if ICNTL(5) = 0, 0 matrix centralized, 1 structure on
* host and mumps give the mapping, 2 structure on host and distributed matrix
* for facto, 3 distributed matrix
*
* ICNTL(19) : > 0, Shur complement returned
*
* ICNTL(20) : 0 rhs dense, 1 rhs sparse
*
* ICNTL(21) : 0 solution in rhs, 1 solution distributed in ISOL_loc and SOL_loc
* allocated by user
*
* ICNTL(22) : 0 in-core, 1 out-of-core
*
* ICNTL(23) : maximum memory allocatable by mumps pre proc
*
* ICNTL(24) : controls the detection of "null pivot rows"
*
* ICNTL(25) :
*
* ICNTL(26) :
*
* ICNTL(27) :
*
* ICNTL(28) : 0 automatic choice, 1 sequential analysis, 2 parallel analysis
*
* ICNTL(29) : 0 automatic choice, 1 PT-Scotch, 2 ParMetis
*/
/* -------------------------------------------------------------------------- */
#include "aka_common.hh"
#include "dof_manager_default.hh"
#include "dof_synchronizer.hh"
#include "sparse_matrix_aij.hh"
#if defined(AKANTU_USE_MPI)
#include "mpi_communicator_data.hh"
#endif
#include "sparse_solver_mumps.hh"
/* -------------------------------------------------------------------------- */
// static std::ostream & operator <<(std::ostream & stream, const DMUMPS_STRUC_C
// & _this) {
// stream << "DMUMPS Data [" << std::endl;
// stream << " + job : " << _this.job << std::endl;
// stream << " + par : " << _this.par << std::endl;
// stream << " + sym : " << _this.sym << std::endl;
// stream << " + comm_fortran : " << _this.comm_fortran << std::endl;
// stream << " + nz : " << _this.nz << std::endl;
// stream << " + irn : " << _this.irn << std::endl;
// stream << " + jcn : " << _this.jcn << std::endl;
// stream << " + nz_loc : " << _this.nz_loc << std::endl;
// stream << " + irn_loc : " << _this.irn_loc << std::endl;
// stream << " + jcn_loc : " << _this.jcn_loc << std::endl;
// stream << "]";
// return stream;
// }
namespace akantu {
/* -------------------------------------------------------------------------- */
SparseSolverMumps::SparseSolverMumps(DOFManagerDefault & dof_manager,
const ID & matrix_id, const ID & id,
const MemoryID & memory_id)
: SparseSolver(dof_manager, matrix_id, id, memory_id),
dof_manager(dof_manager), master_rhs_solution(0, 1) {
AKANTU_DEBUG_IN();
this->prank = communicator.whoAmI();
#ifdef AKANTU_USE_MPI
this->parallel_method = _fully_distributed;
#else // AKANTU_USE_MPI
this->parallel_method = _not_parallel;
#endif // AKANTU_USE_MPI
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
SparseSolverMumps::~SparseSolverMumps() {
AKANTU_DEBUG_IN();
mumpsDataDestroy();
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::mumpsDataDestroy() {
if (this->is_initialized) {
this->mumps_data.job = _smj_destroy; // destroy
dmumps_c(&this->mumps_data);
this->is_initialized = false;
}
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::destroyInternalData() { mumpsDataDestroy(); }
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::checkInitialized() {
if (this->is_initialized)
return;
this->initialize();
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::setOutputLevel() {
// Output setup
icntl(1) = 0; // error output
icntl(2) = 0; // diagnostics output
icntl(3) = 0; // information
icntl(4) = 0;
#if !defined(AKANTU_NDEBUG)
DebugLevel dbg_lvl = debug::debugger.getDebugLevel();
if (AKANTU_DEBUG_TEST(dblDump)) {
strcpy(this->mumps_data.write_problem, "mumps_matrix.mtx");
}
// clang-format off
icntl(1) = (dbg_lvl >= dblWarning) ? 6 : 0;
icntl(3) = (dbg_lvl >= dblInfo) ? 6 : 0;
icntl(2) = (dbg_lvl >= dblTrace) ? 6 : 0;
icntl(4) =
dbg_lvl >= dblDump ? 4 :
dbg_lvl >= dblTrace ? 3 :
dbg_lvl >= dblInfo ? 2 :
dbg_lvl >= dblWarning ? 1 :
0;
// clang-format on
#endif
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::initMumpsData() {
auto & A = dof_manager.getMatrix(matrix_id);
// Default Scaling
icntl(8) = 77;
// Assembled matrix
icntl(5) = 0;
/// Default centralized dense second member
icntl(20) = 0;
icntl(21) = 0;
// automatic choice for analysis
icntl(28) = 0;
UInt size = A.size();
if (prank == 0) {
this->master_rhs_solution.resize(size);
}
this->mumps_data.nz_alloc = 0;
this->mumps_data.n = size;
switch (this->parallel_method) {
case _fully_distributed:
icntl(18) = 3; // fully distributed
this->mumps_data.nz_loc = A.getNbNonZero();
this->mumps_data.irn_loc = A.getIRN().storage();
this->mumps_data.jcn_loc = A.getJCN().storage();
break;
case _not_parallel:
case _master_slave_distributed:
icntl(18) = 0; // centralized
if (prank == 0) {
this->mumps_data.nz = A.getNbNonZero();
this->mumps_data.irn = A.getIRN().storage();
this->mumps_data.jcn = A.getJCN().storage();
} else {
this->mumps_data.nz = 0;
this->mumps_data.irn = nullptr;
this->mumps_data.jcn = nullptr;
}
break;
default:
AKANTU_ERROR("This case should not happen!!");
}
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::initialize() {
AKANTU_DEBUG_IN();
this->mumps_data.par = 1; // The host is part of computations
switch (this->parallel_method) {
case _not_parallel:
break;
case _master_slave_distributed:
this->mumps_data.par = 0; // The host is not part of the computations
/* FALLTHRU */
/* [[fallthrough]]; un-comment when compiler will get it */
case _fully_distributed:
#ifdef AKANTU_USE_MPI
const auto & mpi_data = dynamic_cast<const MPICommunicatorData &>(
communicator.getCommunicatorData());
MPI_Comm mpi_comm = mpi_data.getMPICommunicator();
this->mumps_data.comm_fortran = MPI_Comm_c2f(mpi_comm);
#else
AKANTU_ERROR(
"You cannot use parallel method to solve without activating MPI");
#endif
break;
}
const auto & A = dof_manager.getMatrix(matrix_id);
this->mumps_data.sym = 2 * (A.getMatrixType() == _symmetric);
this->prank = communicator.whoAmI();
this->setOutputLevel();
this->mumps_data.job = _smj_initialize; // initialize
dmumps_c(&this->mumps_data);
this->setOutputLevel();
this->is_initialized = true;
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::analysis() {
AKANTU_DEBUG_IN();
initMumpsData();
this->mumps_data.job = _smj_analyze; // analyze
dmumps_c(&this->mumps_data);
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::factorize() {
AKANTU_DEBUG_IN();
auto & A = dof_manager.getMatrix(matrix_id);
if (parallel_method == _fully_distributed)
this->mumps_data.a_loc = A.getA().storage();
else {
if (prank == 0)
this->mumps_data.a = A.getA().storage();
}
this->mumps_data.job = _smj_factorize; // factorize
dmumps_c(&this->mumps_data);
this->printError();
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::solve(Array<Real> & x, const Array<Real> & b) {
auto & synch = this->dof_manager.getSynchronizer();
if (this->prank == 0) {
this->master_rhs_solution.resize(this->dof_manager.getSystemSize());
synch.gather(b, this->master_rhs_solution);
} else {
synch.gather(b);
}
this->solveInternal();
if (this->prank == 0) {
synch.scatter(x, this->master_rhs_solution);
} else {
synch.scatter(x);
}
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::solve() {
this->master_rhs_solution.copy(this->dof_manager.getGlobalResidual());
this->solveInternal();
this->dof_manager.setGlobalSolution(this->master_rhs_solution);
this->dof_manager.splitSolutionPerDOFs();
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::solveInternal() {
AKANTU_DEBUG_IN();
this->checkInitialized();
const auto & A = dof_manager.getMatrix(matrix_id);
this->setOutputLevel();
if (this->last_profile_release != A.getProfileRelease()) {
this->analysis();
this->last_profile_release = A.getProfileRelease();
}
if (AKANTU_DEBUG_TEST(dblDump)) {
A.saveMatrix("solver_mumps" + std::to_string(prank) + ".mtx");
}
if (this->last_value_release != A.getValueRelease()) {
this->factorize();
this->last_value_release = A.getValueRelease();
}
if (prank == 0) {
this->mumps_data.rhs = this->master_rhs_solution.storage();
}
this->mumps_data.job = _smj_solve; // solve
dmumps_c(&this->mumps_data);
this->printError();
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
void SparseSolverMumps::printError() {
Vector<Int> _info_v(2);
_info_v[0] = info(1); // to get errors
_info_v[1] = -info(1); // to get warnings
dof_manager.getCommunicator().allReduce(_info_v, SynchronizerOperation::_min);
_info_v[1] = -_info_v[1];
if (_info_v[0] < 0) { // < 0 is an error
switch (_info_v[0]) {
case -10: {
AKANTU_CUSTOM_EXCEPTION(
debug::SingularMatrixException(dof_manager.getMatrix(matrix_id)));
break;
}
case -9: {
icntl(14) += 10;
if (icntl(14) != 90) {
// std::cout << "Dynamic memory increase of 10%" << std::endl;
AKANTU_DEBUG_WARNING("MUMPS dynamic memory is insufficient it will be "
"increased allowed to use 10% more");
// change releases to force a recompute
this->last_value_release--;
this->last_profile_release--;
this->solve();
} else {
AKANTU_ERROR("The MUMPS workarray is too small INFO(2)="
<< info(2) << "No further increase possible");
}
break;
}
default:
AKANTU_ERROR("Error in mumps during solve process, check mumps "
"user guide INFO(1) = "
<< _info_v[1]);
}
} else if (_info_v[1] > 0) {
AKANTU_DEBUG_WARNING("Warning in mumps during solve process, check mumps "
"user guide INFO(1) = "
<< _info_v[1]);
}
}
} // akantu
c4science · Help