ex1p.cpp

Go to the documentation of this file.

//                       MFEM Example 1 - Parallel Version
//                              AmgX Modification
//
// Compile with: make ex1p
//
// AmgX sample runs:
//               mpirun -np 4 ex1p
//               mpirun -np 4 ex1p -d cuda
//               mpirun -np 10 ex1p --amgx-file amg_pcg.json --amgx-mpi-teams
//               mpirun -np 4 ex1p --amgx-file amg_pcg.json
//
// Description:  This example code demonstrates the use of MFEM to define a
//               simple finite element discretization of the Laplace problem
//               -Delta u = 1 with homogeneous Dirichlet boundary conditions.
//               Specifically, we discretize using a FE space of the specified
//               order, or if order < 1 using an isoparametric/isogeometric
//               space (i.e. quadratic for quadratic curvilinear mesh, NURBS for
//               NURBS mesh, etc.)
//
//               The example highlights the use of mesh refinement, finite
//               element grid functions, as well as linear and bilinear forms
//               corresponding to the left-hand side and right-hand side of the
//               discrete linear system. We also cover the explicit elimination
//               of essential boundary conditions, static condensation, and the
//               optional connection to the GLVis tool for visualization.
 
#include "mfem.hpp"
#include <fstream>
#include <iostream>
 
using namespace std;
using namespace mfem;
 
#ifndef MFEM_USE_AMGX
#error This example requires that MFEM is built with MFEM_USE_AMGX=YES
#endif
 
int main(int argc, char *argv[])
{
   // 1. Initialize MPI and HYPRE.
   Mpi::Init(argc, argv);
   int num_procs = Mpi::WorldSize();
   int myid = Mpi::WorldRank();
   Hypre::Init();
 
   // 2. Parse command-line options.
   const char *mesh_file = "../../data/star.mesh";
   int order = 1;
   bool static_cond = false;
   bool pa = false;
   const char *device_config = "cpu";
   bool visualization = true;
   bool amgx_lib = true;
   bool amgx_mpi_teams = false;
   const char* amgx_json_file = ""; // JSON file for AmgX
   int ndevices = 1;
 
   OptionsParser args(argc, argv);
   args.AddOption(&mesh_file, "-m", "--mesh",
                  "Mesh file to use.");
   args.AddOption(&order, "-o", "--order",
                  "Finite element order (polynomial degree) or -1 for"
                  " isoparametric space.");
   args.AddOption(&static_cond, "-sc", "--static-condensation", "-no-sc",
                  "--no-static-condensation", "Enable static condensation.");
   args.AddOption(&pa, "-pa", "--partial-assembly", "-no-pa",
                  "--no-partial-assembly", "Enable Partial Assembly.");
   args.AddOption(&amgx_lib, "-amgx", "--amgx-lib", "-no-amgx",
                  "--no-amgx-lib", "Use AmgX in example.");
   args.AddOption(&amgx_json_file, "--amgx-file", "--amgx-file",
                  "AMGX solver config file (overrides --amgx-solver, --amgx-verbose)");
   args.AddOption(&amgx_mpi_teams, "--amgx-mpi-teams", "--amgx-mpi-teams",
                  "--amgx-mpi-gpu-exclusive", "--amgx-mpi-gpu-exclusive",
                  "Create MPI teams when using AmgX to load balance between ranks and GPUs.");
   args.AddOption(&device_config, "-d", "--device",
                  "Device configuration string, see Device::Configure().");
   args.AddOption(&visualization, "-vis", "--visualization", "-no-vis",
                  "--no-visualization",
                  "Enable or disable GLVis visualization.");
   args.AddOption(&ndevices, "-nd","--gpus-per-node-in-teams-mode",
                  "Number of GPU devices per node (Only used if amgx_mpi_teams is true).");
 
   args.Parse();
   if (!args.Good())
   {
      if (myid == 0)
      {
         args.PrintUsage(cout);
      }
      return 1;
   }
   if (myid == 0)
   {
      args.PrintOptions(cout);
   }
 
   // 3. Enable hardware devices such as GPUs, and programming models such as
   //    CUDA, OCCA, RAJA and OpenMP based on command line options.
   Device device(device_config);
   if (myid == 0) { device.Print(); }
 
   // 4. Read the (serial) mesh from the given mesh file on all processors.  We
   //    can handle triangular, quadrilateral, tetrahedral, hexahedral, surface
   //    and volume meshes with the same code.
   Mesh mesh(mesh_file, 1, 1);
   int dim = mesh.Dimension();
 
   // 5. Refine the serial mesh on all processors to increase the resolution. In
   //    this example we do 'ref_levels' of uniform refinement. We choose
   //    'ref_levels' to be the largest number that gives a final mesh with no
   //    more than 10,000 elements.
   {
      int ref_levels =
         (int)floor(log(10000./mesh.GetNE())/log(2.)/dim);
      for (int l = 0; l < ref_levels; l++)
      {
         mesh.UniformRefinement();
      }
   }
 
   // 6. Define a parallel mesh by a partitioning of the serial mesh. Refine
   //    this mesh further in parallel to increase the resolution. Once the
   //    parallel mesh is defined, the serial mesh can be deleted.
   ParMesh pmesh(MPI_COMM_WORLD, mesh);
   mesh.Clear();
   {
      int par_ref_levels = 2;
      for (int l = 0; l < par_ref_levels; l++)
      {
         pmesh.UniformRefinement();
      }
   }
 
   // 7. Define a parallel finite element space on the parallel mesh. Here we
   //    use continuous Lagrange finite elements of the specified order. If
   //    order < 1, we instead use an isoparametric/isogeometric space.
   FiniteElementCollection *fec;
   bool delete_fec;
   if (order > 0)
   {
      fec = new H1_FECollection(order, dim);
      delete_fec = true;
   }
   else if (pmesh.GetNodes())
   {
      fec = pmesh.GetNodes()->OwnFEC();
      delete_fec = false;
      if (myid == 0)
      {
         cout << "Using isoparametric FEs: " << fec->Name() << endl;
      }
   }
   else
   {
      fec = new H1_FECollection(order = 1, dim);
      delete_fec = true;
   }
   ParFiniteElementSpace fespace(&pmesh, fec);
   HYPRE_BigInt size = fespace.GlobalTrueVSize();
   if (myid == 0)
   {
      cout << "Number of finite element unknowns: " << size << endl;
   }
 
   // 8. Determine the list of true (i.e. parallel conforming) essential
   //    boundary dofs. In this example, the boundary conditions are defined
   //    by marking all the boundary attributes from the mesh as essential
   //    (Dirichlet) and converting them to a list of true dofs.
   Array<int> ess_tdof_list;
   if (pmesh.bdr_attributes.Size())
   {
      Array<int> ess_bdr(pmesh.bdr_attributes.Max());
      ess_bdr = 1;
      fespace.GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
   }
 
   // 9. Set up the parallel linear form b(.) which corresponds to the
   //    right-hand side of the FEM linear system, which in this case is
   //    (1,phi_i) where phi_i are the basis functions in fespace.
   ParLinearForm b(&fespace);
   ConstantCoefficient one(1.0);
   b.AddDomainIntegrator(new DomainLFIntegrator(one));
   b.Assemble();
 
   // 10. Define the solution vector x as a parallel finite element grid function
   //     corresponding to fespace. Initialize x with initial guess of zero,
   //     which satisfies the boundary conditions.
   ParGridFunction x(&fespace);
   x = 0.0;
 
   // 11. Set up the parallel bilinear form a(.,.) on the finite element space
   //     corresponding to the Laplacian operator -Delta, by adding the Diffusion
   //     domain integrator.
   ParBilinearForm a(&fespace);
   if (pa) { a.SetAssemblyLevel(AssemblyLevel::PARTIAL); }
   a.AddDomainIntegrator(new DiffusionIntegrator(one));
 
   // 12. Assemble the parallel bilinear form and the corresponding linear
   //     system, applying any necessary transformations such as: parallel
   //     assembly, eliminating boundary conditions, applying conforming
   //     constraints for non-conforming AMR, static condensation, etc.
   if (static_cond) { a.EnableStaticCondensation(); }
   a.Assemble();
 
   OperatorPtr A;
   Vector B, X;
   a.FormLinearSystem(ess_tdof_list, x, b, A, X, B);
 
   // 13. Solve the linear system A X = B.
   //     * With full assembly, use the BoomerAMG preconditioner from hypre.
   //     * If AmgX is available solve using amg preconditioner.
   //     * With partial assembly, use Jacobi smoothing, for now.
   Solver *prec = NULL;
   if (pa)
   {
      if (UsesTensorBasis(fespace))
      {
         prec = new OperatorJacobiSmoother(a, ess_tdof_list);
      }
 
      CGSolver cg(MPI_COMM_WORLD);
      cg.SetRelTol(1e-12);
      cg.SetMaxIter(2000);
      cg.SetPrintLevel(1);
      if (prec) { cg.SetPreconditioner(*prec); }
      cg.SetOperator(*A);
      cg.Mult(B, X);
      delete prec;
   }
   else if (amgx_lib && strcmp(amgx_json_file,"") == 0)
   {
      MFEM_VERIFY(!amgx_mpi_teams,
                  "Please add JSON file to try AmgX with MPI teams mode");
 
      bool amgx_verbose = false;
      prec = new AmgXSolver(MPI_COMM_WORLD, AmgXSolver::PRECONDITIONER,
                            amgx_verbose);
 
      CGSolver cg(MPI_COMM_WORLD);
      cg.SetRelTol(1e-12);
      cg.SetMaxIter(2000);
      cg.SetPrintLevel(1);
      if (prec) { cg.SetPreconditioner(*prec); }
      cg.SetOperator(*A);
      cg.Mult(B, X);
      delete prec;
 
   }
   else if (amgx_lib && strcmp(amgx_json_file,"") != 0)
   {
      AmgXSolver amgx;
      amgx.ReadParameters(amgx_json_file, AmgXSolver::EXTERNAL);
 
      if (amgx_mpi_teams)
      {
         // Forms MPI teams to load balance between MPI ranks and GPUs
         amgx.InitMPITeams(MPI_COMM_WORLD, ndevices);
      }
      else
      {
         // Assumes each MPI rank is paired with a GPU
         amgx.InitExclusiveGPU(MPI_COMM_WORLD);
      }
 
      amgx.SetOperator(*A.As<HypreParMatrix>());
      amgx.SetConvergenceCheck(true);
      amgx.Mult(B, X);
 
      // Release MPI communicators and resources created by AmgX
      amgx.Finalize();
   }
   else
   {
      prec = new HypreBoomerAMG;
 
      CGSolver cg(MPI_COMM_WORLD);
      cg.SetRelTol(1e-12);
      cg.SetMaxIter(2000);
      cg.SetPrintLevel(1);
      if (prec) { cg.SetPreconditioner(*prec); }
      cg.SetOperator(*A);
      cg.Mult(B, X);
      delete prec;
   }
 
   // 14. Recover the parallel grid function corresponding to X. This is the
   //     local finite element solution on each processor.
   a.RecoverFEMSolution(X, b, x);
 
   // 15. Save the refined mesh and the solution in parallel. This output can
   //     be viewed later using GLVis: "glvis -np <np> -m mesh -g sol".
   {
      ostringstream mesh_name, sol_name;
      mesh_name << "mesh." << setfill('0') << setw(6) << myid;
      sol_name << "sol." << setfill('0') << setw(6) << myid;
 
      ofstream mesh_ofs(mesh_name.str().c_str());
      mesh_ofs.precision(8);
      pmesh.Print(mesh_ofs);
 
      ofstream sol_ofs(sol_name.str().c_str());
      sol_ofs.precision(8);
      x.Save(sol_ofs);
   }
 
   // 16. Send the solution by socket to a GLVis server.
   if (visualization)
   {
      char vishost[] = "localhost";
      int  visport   = 19916;
      socketstream sol_sock(vishost, visport);
      sol_sock << "parallel " << num_procs << " " << myid << "\n";
      sol_sock.precision(8);
      sol_sock << "solution\n" << pmesh << x << flush;
   }
 
   // 17. Free the used memory.
   if (delete_fec)
   {
      delete fec;
   }
 
   return 0;
}