Perf fenl(
MPI_Comm comm ,
const int use_print ,
const int use_trials ,
const int use_atomic ,
const int use_elems[] )
{
typedef Kokkos::Example::BoxElemFixture< Space , ElemOrder > FixtureType ;
typedef Kokkos::Example::CrsMatrix< double , Space >
SparseMatrixType ;
typedef typename SparseMatrixType::StaticCrsGraphType
SparseGraphType ;
typedef Kokkos::Example::FENL::NodeNodeGraph< typename FixtureType::elem_node_type , SparseGraphType , FixtureType::ElemNode >
NodeNodeGraphType ;
typedef Kokkos::Example::FENL::ElementComputation< FixtureType , SparseMatrixType >
ElementComputationType ;
typedef Kokkos::Example::FENL::DirichletComputation< FixtureType , SparseMatrixType >
DirichletComputationType ;
typedef NodeElemGatherFill< ElementComputationType >
NodeElemGatherFillType ;
typedef typename ElementComputationType::vector_type VectorType ;
typedef Kokkos::Example::VectorImport<
typename FixtureType::comm_list_type ,
typename FixtureType::send_nodeid_type ,
VectorType > ImportType ;
//------------------------------------
const unsigned newton_iteration_limit = 10 ;
const double newton_iteration_tolerance = 1e-7 ;
const unsigned cg_iteration_limit = 200 ;
const double cg_iteration_tolerance = 1e-7 ;
//------------------------------------
const int print_flag = use_print && std::is_same< Kokkos::HostSpace , typename Space::memory_space >::value ;
int comm_rank ;
int comm_size ;
MPI_Comm_rank( comm , & comm_rank );
MPI_Comm_size( comm , & comm_size );
// Decompose by node to avoid mpi-communication for assembly
const float bubble_x = 1.0 ;
const float bubble_y = 1.0 ;
const float bubble_z = 1.0 ;
const FixtureType fixture( BoxElemPart::DecomposeNode , comm_size , comm_rank ,
use_elems[0] , use_elems[1] , use_elems[2] ,
bubble_x , bubble_y , bubble_z );
{
int global_error = ! fixture.ok();
#if defined( KOKKOS_ENABLE_MPI )
int local_error = global_error ;
global_error = 0 ;
MPI_Allreduce( & local_error , & global_error , 1 , MPI_INT , MPI_SUM , comm );
#endif
if ( global_error ) {
throw std::runtime_error(std::string("Error generating finite element fixture"));
}
}
//------------------------------------
const ImportType comm_nodal_import(
comm ,
fixture.recv_node() ,
fixture.send_node() ,
fixture.send_nodeid() ,
fixture.node_count_owned() ,
fixture.node_count() - fixture.node_count_owned() );
//------------------------------------
const double bc_lower_value = 1 ;
const double bc_upper_value = 2 ;
const Kokkos::Example::FENL::ManufacturedSolution
manufactured_solution( 0 , 1 , bc_lower_value , bc_upper_value );
//------------------------------------
for ( int k = 0 ; k < comm_size && use_print ; ++k ) {
if ( k == comm_rank ) {
typename FixtureType::node_grid_type::HostMirror
h_node_grid = Kokkos::create_mirror_view( fixture.node_grid() );
typename FixtureType::node_coord_type::HostMirror
h_node_coord = Kokkos::create_mirror_view( fixture.node_coord() );
typename FixtureType::elem_node_type::HostMirror
h_elem_node = Kokkos::create_mirror_view( fixture.elem_node() );
Kokkos::deep_copy( h_node_grid , fixture.node_grid() );
Kokkos::deep_copy( h_node_coord , fixture.node_coord() );
Kokkos::deep_copy( h_elem_node , fixture.elem_node() );
std::cout << "MPI[" << comm_rank << "]" << std::endl ;
std::cout << "Node grid {" ;
for ( unsigned inode = 0 ; inode < fixture.node_count() ; ++inode ) {
std::cout << " (" << h_node_grid(inode,0)
<< "," << h_node_grid(inode,1)
<< "," << h_node_grid(inode,2)
<< ")" ;
}
std::cout << " }" << std::endl ;
std::cout << "Node coord {" ;
for ( unsigned inode = 0 ; inode < fixture.node_count() ; ++inode ) {
std::cout << " (" << h_node_coord(inode,0)
<< "," << h_node_coord(inode,1)
<< "," << h_node_coord(inode,2)
<< ")" ;
}
std::cout << " }" << std::endl ;
std::cout << "Manufactured solution"
<< " a[" << manufactured_solution.a << "]"
<< " b[" << manufactured_solution.b << "]"
<< " K[" << manufactured_solution.K << "]"
<< " {" ;
for ( unsigned inode = 0 ; inode < fixture.node_count() ; ++inode ) {
std::cout << " " << manufactured_solution( h_node_coord( inode , 2 ) );
}
std::cout << " }" << std::endl ;
std::cout << "ElemNode {" << std::endl ;
for ( unsigned ielem = 0 ; ielem < fixture.elem_count() ; ++ielem ) {
std::cout << " elem[" << ielem << "]{" ;
for ( unsigned inode = 0 ; inode < FixtureType::ElemNode ; ++inode ) {
std::cout << " " << h_elem_node(ielem,inode);
}
std::cout << " }{" ;
for ( unsigned inode = 0 ; inode < FixtureType::ElemNode ; ++inode ) {
std::cout << " (" << h_node_grid(h_elem_node(ielem,inode),0)
<< "," << h_node_grid(h_elem_node(ielem,inode),1)
<< "," << h_node_grid(h_elem_node(ielem,inode),2)
<< ")" ;
}
std::cout << " }" << std::endl ;
}
std::cout << "}" << std::endl ;
}
std::cout.flush();
MPI_Barrier( comm );
}
//------------------------------------
Kokkos::Timer wall_clock ;
Perf perf_stats = Perf() ;
for ( int itrial = 0 ; itrial < use_trials ; ++itrial ) {
Perf perf = Perf() ;
perf.global_elem_count = fixture.elem_count_global();
perf.global_node_count = fixture.node_count_global();
//----------------------------------
// Create the sparse matrix graph and element-to-graph map
// from the element->to->node identifier array.
// The graph only has rows for the owned nodes.
typename NodeNodeGraphType::Times graph_times;
const NodeNodeGraphType
mesh_to_graph( fixture.elem_node() , fixture.node_count_owned(), graph_times );
perf.map_ratio = maximum(comm, graph_times.ratio);
perf.fill_node_set = maximum(comm, graph_times.fill_node_set);
perf.scan_node_count = maximum(comm, graph_times.scan_node_count);
perf.fill_graph_entries = maximum(comm, graph_times.fill_graph_entries);
perf.sort_graph_entries = maximum(comm, graph_times.sort_graph_entries);
perf.fill_element_graph = maximum(comm, graph_times.fill_element_graph);
wall_clock.reset();
// Create the sparse matrix from the graph:
SparseMatrixType jacobian( mesh_to_graph.graph );
Space::fence();
perf.create_sparse_matrix = maximum( comm , wall_clock.seconds() );
//----------------------------------
for ( int k = 0 ; k < comm_size && print_flag ; ++k ) {
if ( k == comm_rank ) {
const unsigned nrow = jacobian.graph.numRows();
std::cout << "MPI[" << comm_rank << "]" << std::endl ;
std::cout << "JacobianGraph {" << std::endl ;
for ( unsigned irow = 0 ; irow < nrow ; ++irow ) {
std::cout << " row[" << irow << "]{" ;
const unsigned entry_end = jacobian.graph.row_map(irow+1);
for ( unsigned entry = jacobian.graph.row_map(irow) ; entry < entry_end ; ++entry ) {
std::cout << " " << jacobian.graph.entries(entry);
}
std::cout << " }" << std::endl ;
}
std::cout << "}" << std::endl ;
std::cout << "ElemGraph {" << std::endl ;
for ( unsigned ielem = 0 ; ielem < mesh_to_graph.elem_graph.dimension_0() ; ++ielem ) {
std::cout << " elem[" << ielem << "]{" ;
for ( unsigned irow = 0 ; irow < mesh_to_graph.elem_graph.dimension_1() ; ++irow ) {
std::cout << " {" ;
for ( unsigned icol = 0 ; icol < mesh_to_graph.elem_graph.dimension_2() ; ++icol ) {
std::cout << " " << mesh_to_graph.elem_graph(ielem,irow,icol);
}
std::cout << " }" ;
}
std::cout << " }" << std::endl ;
}
std::cout << "}" << std::endl ;
}
std::cout.flush();
MPI_Barrier( comm );
}
//----------------------------------
// Allocate solution vector for each node in the mesh and residual vector for each owned node
const VectorType nodal_solution( "nodal_solution" , fixture.node_count() );
const VectorType nodal_residual( "nodal_residual" , fixture.node_count_owned() );
const VectorType nodal_delta( "nodal_delta" , fixture.node_count_owned() );
// Create element computation functor
const ElementComputationType elemcomp(
use_atomic ? ElementComputationType( fixture , manufactured_solution.K , nodal_solution ,
mesh_to_graph.elem_graph , jacobian , nodal_residual )
: ElementComputationType( fixture , manufactured_solution.K , nodal_solution ) );
const NodeElemGatherFillType gatherfill(
use_atomic ? NodeElemGatherFillType()
: NodeElemGatherFillType( fixture.elem_node() ,
mesh_to_graph.elem_graph ,
nodal_residual ,
jacobian ,
elemcomp.elem_residuals ,
elemcomp.elem_jacobians ) );
// Create boundary condition functor
const DirichletComputationType dirichlet(
fixture , nodal_solution , jacobian , nodal_residual ,
2 /* apply at 'z' ends */ ,
manufactured_solution.T_zmin ,
manufactured_solution.T_zmax );
//----------------------------------
// Nonlinear Newton iteration:
double residual_norm_init = 0 ;
for ( perf.newton_iter_count = 0 ;
perf.newton_iter_count < newton_iteration_limit ;
++perf.newton_iter_count ) {
//--------------------------------
comm_nodal_import( nodal_solution );
//--------------------------------
// Element contributions to residual and jacobian
wall_clock.reset();
Kokkos::deep_copy( nodal_residual , double(0) );
Kokkos::deep_copy( jacobian.coeff , double(0) );
elemcomp.apply();
if ( ! use_atomic ) {
gatherfill.apply();
}
Space::fence();
perf.fill_time = maximum( comm , wall_clock.seconds() );
//--------------------------------
// Apply boundary conditions
wall_clock.reset();
dirichlet.apply();
Space::fence();
perf.bc_time = maximum( comm , wall_clock.seconds() );
//--------------------------------
// Evaluate convergence
const double residual_norm =
std::sqrt(
Kokkos::Example::all_reduce(
Kokkos::Example::dot( fixture.node_count_owned() , nodal_residual, nodal_residual ) , comm ) );
perf.newton_residual = residual_norm ;
if ( 0 == perf.newton_iter_count ) { residual_norm_init = residual_norm ; }
if ( residual_norm < residual_norm_init * newton_iteration_tolerance ) { break ; }
//--------------------------------
// Solve for nonlinear update
CGSolveResult cg_result ;
Kokkos::Example::cgsolve( comm_nodal_import
, jacobian
, nodal_residual
, nodal_delta
, cg_iteration_limit
, cg_iteration_tolerance
, & cg_result
);
// Update solution vector
Kokkos::Example::waxpby( fixture.node_count_owned() , nodal_solution , -1.0 , nodal_delta , 1.0 , nodal_solution );
perf.cg_iter_count += cg_result.iteration ;
perf.matvec_time += cg_result.matvec_time ;
perf.cg_time += cg_result.iter_time ;
//--------------------------------
if ( print_flag ) {
const double delta_norm =
std::sqrt(
Kokkos::Example::all_reduce(
Kokkos::Example::dot( fixture.node_count_owned() , nodal_delta, nodal_delta ) , comm ) );
if ( 0 == comm_rank ) {
std::cout << "Newton iteration[" << perf.newton_iter_count << "]"
<< " residual[" << perf.newton_residual << "]"
<< " update[" << delta_norm << "]"
<< " cg_iteration[" << cg_result.iteration << "]"
<< " cg_residual[" << cg_result.norm_res << "]"
<< std::endl ;
}
for ( int k = 0 ; k < comm_size ; ++k ) {
if ( k == comm_rank ) {
const unsigned nrow = jacobian.graph.numRows();
std::cout << "MPI[" << comm_rank << "]" << std::endl ;
std::cout << "Residual {" ;
for ( unsigned irow = 0 ; irow < nrow ; ++irow ) {
std::cout << " " << nodal_residual(irow);
}
std::cout << " }" << std::endl ;
std::cout << "Delta {" ;
for ( unsigned irow = 0 ; irow < nrow ; ++irow ) {
std::cout << " " << nodal_delta(irow);
}
std::cout << " }" << std::endl ;
std::cout << "Solution {" ;
for ( unsigned irow = 0 ; irow < nrow ; ++irow ) {
std::cout << " " << nodal_solution(irow);
}
std::cout << " }" << std::endl ;
std::cout << "Jacobian[ "
<< jacobian.graph.numRows() << " x " << Kokkos::maximum_entry( jacobian.graph )
<< " ] {" << std::endl ;
for ( unsigned irow = 0 ; irow < nrow ; ++irow ) {
std::cout << " {" ;
const unsigned entry_end = jacobian.graph.row_map(irow+1);
for ( unsigned entry = jacobian.graph.row_map(irow) ; entry < entry_end ; ++entry ) {
std::cout << " (" << jacobian.graph.entries(entry)
<< "," << jacobian.coeff(entry)
<< ")" ;
}
std::cout << " }" << std::endl ;
}
std::cout << "}" << std::endl ;
}
std::cout.flush();
MPI_Barrier( comm );
}
}
//--------------------------------
}
// Evaluate solution error
if ( 0 == itrial ) {
const typename FixtureType::node_coord_type::HostMirror
h_node_coord = Kokkos::create_mirror_view( fixture.node_coord() );
const typename VectorType::HostMirror
h_nodal_solution = Kokkos::create_mirror_view( nodal_solution );
Kokkos::deep_copy( h_node_coord , fixture.node_coord() );
Kokkos::deep_copy( h_nodal_solution , nodal_solution );
double error_max = 0 ;
for ( unsigned inode = 0 ; inode < fixture.node_count_owned() ; ++inode ) {
const double answer = manufactured_solution( h_node_coord( inode , 2 ) );
const double error = ( h_nodal_solution(inode) - answer ) / answer ;
if ( error_max < fabs( error ) ) { error_max = fabs( error ); }
}
perf.error_max = std::sqrt( Kokkos::Example::all_reduce_max( error_max , comm ) );
perf_stats = perf ;
}
else {
perf_stats.fill_node_set = std::min( perf_stats.fill_node_set , perf.fill_node_set );
perf_stats.scan_node_count = std::min( perf_stats.scan_node_count , perf.scan_node_count );
perf_stats.fill_graph_entries = std::min( perf_stats.fill_graph_entries , perf.fill_graph_entries );
perf_stats.sort_graph_entries = std::min( perf_stats.sort_graph_entries , perf.sort_graph_entries );
perf_stats.fill_element_graph = std::min( perf_stats.fill_element_graph , perf.fill_element_graph );
perf_stats.create_sparse_matrix = std::min( perf_stats.create_sparse_matrix , perf.create_sparse_matrix );
perf_stats.fill_time = std::min( perf_stats.fill_time , perf.fill_time );
perf_stats.bc_time = std::min( perf_stats.bc_time , perf.bc_time );
perf_stats.cg_time = std::min( perf_stats.cg_time , perf.cg_time );
}
}
return perf_stats ;
}
Perf fenl_assembly(
const Teuchos::RCP<const Teuchos::Comm<int> >& comm ,
const int use_print ,
const int use_trials ,
const int use_atomic ,
const int use_nodes[] ,
Kokkos::DeviceConfig dev_config ,
Kokkos::View< Scalar* , Kokkos::LayoutLeft, Device >& nodal_residual)
{
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcpFromRef;
using Teuchos::arrayView;
using Teuchos::ParameterList;
typedef Kokkos::Example::BoxElemFixture< Device , Kokkos::Example::BoxElemPart::ElemLinear > FixtureType ;
typedef Kokkos::CrsMatrix< Scalar , unsigned , Device >
//typedef typename GlobalMatrixType::k_local_matrix_type
LocalMatrixType ;
typedef typename LocalMatrixType::StaticCrsGraphType
LocalGraphType ;
typedef Kokkos::Example::FENL::NodeNodeGraph< typename FixtureType::elem_node_type , LocalGraphType , FixtureType::ElemNode >
NodeNodeGraphType ;
typedef Kokkos::Example::FENL::ElementComputationConstantCoefficient CoeffFunctionType;
typedef Kokkos::Example::FENL::ElementComputation< FixtureType , LocalMatrixType , CoeffFunctionType >
ElementComputationType ;
typedef Kokkos::Example::FENL::DirichletComputation< FixtureType , LocalMatrixType >
DirichletComputationType ;
typedef Kokkos::Example::FENL::NodeElemGatherFill< ElementComputationType >
NodeElemGatherFillType ;
typedef typename ElementComputationType::vector_type VectorType ;
typedef Kokkos::Example::VectorImport<
typename FixtureType::comm_list_type ,
typename FixtureType::send_nodeid_type ,
VectorType > ImportType ;
//------------------------------------
const int print_flag = use_print && Kokkos::Impl::is_same< Kokkos::HostSpace , typename Device::memory_space >::value ;
const int comm_rank = comm->getRank();
const int comm_size = comm->getSize();
// Decompose by node to avoid parallel communication in assembly
const float bubble_x = 1.0 ;
const float bubble_y = 1.0 ;
const float bubble_z = 1.0 ;
const FixtureType fixture( Kokkos::Example::BoxElemPart::DecomposeNode ,
comm_size , comm_rank ,
use_nodes[0] , use_nodes[1] , use_nodes[2] ,
bubble_x , bubble_y , bubble_z );
//------------------------------------
const ImportType comm_nodal_import(
comm ,
fixture.recv_node() ,
fixture.send_node() ,
fixture.send_nodeid() ,
fixture.node_count_owned() ,
fixture.node_count() - fixture.node_count_owned() );
//------------------------------------
const double bc_lower_value = 1 ;
const double bc_upper_value = 2 ;
CoeffFunctionType diffusion_coefficient( 1.0 );
//------------------------------------
if ( print_flag ) {
std::cout << "ElemNode {" << std::endl ;
for ( unsigned ielem = 0 ; ielem < fixture.elem_count() ; ++ielem ) {
std::cout << " elem[" << ielem << "]{" ;
for ( unsigned inode = 0 ; inode < FixtureType::ElemNode ; ++inode ) {
std::cout << " " << fixture.elem_node(ielem,inode);
}
std::cout << " }" << std::endl ;
}
std::cout << "}" << std::endl ;
}
//------------------------------------
Kokkos::Impl::Timer wall_clock ;
Perf perf_stats = Perf() ;
for ( int itrial = 0 ; itrial < use_trials ; ++itrial ) {
Perf perf = Perf() ;
perf.global_elem_count = fixture.elem_count_global();
perf.global_node_count = fixture.node_count_global();
//----------------------------------
// Create the local sparse matrix graph and element-to-graph map
// from the element->to->node identifier array.
// The graph only has rows for the owned nodes.
typename NodeNodeGraphType::Times graph_times;
const NodeNodeGraphType
mesh_to_graph( fixture.elem_node() , fixture.node_count_owned(),
graph_times );
// Create the local sparse matrix from the graph:
LocalMatrixType jacobian( "jacobian" , mesh_to_graph.graph );
//----------------------------------
if ( print_flag ) {
const unsigned nrow = jacobian.numRows();
std::cout << "JacobianGraph[ "
<< jacobian.numRows() << " x " << jacobian.numCols()
<< " ] {" << std::endl ;
for ( unsigned irow = 0 ; irow < nrow ; ++irow ) {
std::cout << " row[" << irow << "]{" ;
const unsigned entry_end = jacobian.graph.row_map(irow+1);
for ( unsigned entry = jacobian.graph.row_map(irow) ; entry < entry_end ; ++entry ) {
std::cout << " " << jacobian.graph.entries(entry);
}
std::cout << " }" << std::endl ;
}
std::cout << "}" << std::endl ;
std::cout << "ElemGraph {" << std::endl ;
for ( unsigned ielem = 0 ; ielem < mesh_to_graph.elem_graph.dimension_0() ; ++ielem ) {
std::cout << " elem[" << ielem << "]{" ;
for ( unsigned irow = 0 ; irow < mesh_to_graph.elem_graph.dimension_1() ; ++irow ) {
std::cout << " {" ;
for ( unsigned icol = 0 ; icol < mesh_to_graph.elem_graph.dimension_2() ; ++icol ) {
std::cout << " " << mesh_to_graph.elem_graph(ielem,irow,icol);
}
std::cout << " }" ;
}
std::cout << " }" << std::endl ;
}
std::cout << "}" << std::endl ;
}
//----------------------------------
// Allocate solution vector for each node in the mesh and residual vector for each owned node
VectorType nodal_solution( "nodal_solution" , fixture.node_count() );
nodal_residual = VectorType( "nodal_residual" , fixture.node_count_owned() );
// Get DeviceConfig structs used by some functors
Kokkos::DeviceConfig dev_config_elem, dev_config_gath, dev_config_bc;
Kokkos::Example::FENL::CreateDeviceConfigs<Scalar>::eval( dev_config_elem,
dev_config_gath,
dev_config_bc );
// Create element computation functor
const ElementComputationType elemcomp(
use_atomic ? ElementComputationType( fixture , diffusion_coefficient ,
nodal_solution ,
mesh_to_graph.elem_graph ,
jacobian , nodal_residual ,
dev_config_elem )
: ElementComputationType( fixture , diffusion_coefficient ,
nodal_solution , dev_config_elem ) );
const NodeElemGatherFillType gatherfill(
use_atomic ? NodeElemGatherFillType()
: NodeElemGatherFillType( fixture.elem_node() ,
mesh_to_graph.elem_graph ,
nodal_residual ,
jacobian ,
elemcomp.elem_residuals ,
elemcomp.elem_jacobians ,
dev_config_gath) );
// Create boundary condition functor
const DirichletComputationType dirichlet(
fixture , nodal_solution , jacobian , nodal_residual ,
2 /* apply at 'z' ends */ ,
bc_lower_value ,
bc_upper_value ,
dev_config_bc );
Kokkos::deep_copy( nodal_solution , Scalar(1) );
//--------------------------------
wall_clock.reset();
comm_nodal_import( nodal_solution );
Device::fence();
perf.import_time = maximum( comm , wall_clock.seconds() );
//--------------------------------
// Element contributions to residual and jacobian
wall_clock.reset();
Kokkos::deep_copy( nodal_residual , Scalar(0) );
Kokkos::deep_copy( jacobian.values , Scalar(0) );
elemcomp.apply();
if ( ! use_atomic ) {
gatherfill.apply();
}
Device::fence();
perf.fill_time = maximum( comm , wall_clock.seconds() );
//--------------------------------
// Apply boundary conditions
wall_clock.reset();
dirichlet.apply();
Device::fence();
perf.bc_time = maximum( comm , wall_clock.seconds() );
//--------------------------------
perf_stats.increment(perf);
}
return perf_stats ;
}