C++ Comm::exchange方法代码示例

int main(int argc, char **argv)
{
  //Common miniMD settings
  In in;
  in.datafile = NULL;
  int me=0;                     //local MPI rank
  int nprocs=1;                 //number of MPI ranks
  int num_threads=32;		    //number of Threads per Block threads
  int num_steps=-1;             //number of timesteps (if -1 use value from lj.in)
  int system_size=-1;           //size of the system (if -1 use value from lj.in)
  int check_safeexchange=0;     //if 1 complain if atom moves further than 1 subdomain length between exchanges
  int do_safeexchange=0;        //if 1 use safe exchange mode [allows exchange over multiple subdomains]
  int use_sse=0;                //setting for SSE variant of miniMD only
  int screen_yaml=0;            //print yaml output to screen also
  int yaml_output=0;            //print yaml output
  int halfneigh=0;              //1: use half neighborlist; 0: use full neighborlist; -1: use original miniMD version half neighborlist force
  char* input_file = NULL;
  int ghost_newton = 0;
  int skip_gpu = 999;
  int neighbor_size = -1;

  //OpenCL specific
  int platform = 0;
  int device = 0;
  int subdevice = -1;
  int ppn = 2;
  int use_tex = 0;
  int threads_per_atom = 1;
  int map_device=0;

  for(int i = 0; i < argc; i++) {
    if((strcmp(argv[i], "-i") == 0) || (strcmp(argv[i], "--input_file") == 0)) {
      input_file = argv[++i];
      continue;
    }
    if((strcmp(argv[i],"-p")==0)||(strcmp(argv[i],"--platform")==0)) {platform=atoi(argv[++i]); continue;}
    if((strcmp(argv[i],"-d")==0)||(strcmp(argv[i],"--device")==0)) {device=atoi(argv[++i]); continue;}
	if((strcmp(argv[i],"-sd")==0)||(strcmp(argv[i],"--subdevice")==0)) {subdevice=atoi(argv[++i]); continue;}
	if((strcmp(argv[i],"-sd_map")==0)||(strcmp(argv[i],"--subdevice_mapping")==0)) {subdevice=1-me%ppn; continue;}
	if((strcmp(argv[i],"-ng")==0)||(strcmp(argv[i],"--num_gpus")==0)) {ppn=atoi(argv[++i]); continue;}
	if((strcmp(argv[i],"-dm")==0)||(strcmp(argv[i],"--device_map")==0)) {map_device=1; continue;}
  }

  MPI_Init(&argc, &argv);
  MPI_Comm_rank(MPI_COMM_WORLD, &me);
  MPI_Comm_size(MPI_COMM_WORLD, &nprocs);

  if(map_device) {device = me%ppn; if(device>=skip_gpu) device++;}

  OpenCLWrapper* opencl = new OpenCLWrapper;
  if( me == 0)
  printf("# Platforms: %i\n",opencl->num_platforms);
  printf("# Proc: %i using device %i\n",me,device);
  opencl->Init(argc,argv,device,device+1,NULL,platform,subdevice);

  int error = 0;
  if(input_file == NULL)
    error = input(in, "in.lj.miniMD");
  else
	error = input(in, input_file);

  if (error)
  {
	  MPI_Finalize();
	  exit(0);
  }

  for(int i=0;i<argc;i++)
  {
     if((strcmp(argv[i],"-t")==0)||(strcmp(argv[i],"--num_threads")==0)) {num_threads=atoi(argv[++i]); continue;}
     if((strcmp(argv[i],"-n")==0)||(strcmp(argv[i],"--nsteps")==0))  {num_steps=atoi(argv[++i]); continue;}
     if((strcmp(argv[i],"-s")==0)||(strcmp(argv[i],"--size")==0))  {system_size=atoi(argv[++i]); continue;}
     if((strcmp(argv[i],"--half_neigh")==0))  {halfneigh=atoi(argv[++i]); continue;}
     if((strcmp(argv[i],"-sse")==0))  {use_sse=atoi(argv[++i]); continue;}
     if((strcmp(argv[i],"--check_exchange")==0))  {check_safeexchange=1; continue;}
     if((strcmp(argv[i],"-o")==0)||(strcmp(argv[i],"--yaml_output")==0))  {yaml_output=atoi(argv[++i]); continue;}
     if((strcmp(argv[i],"--yaml_screen")==0))  {screen_yaml=1; continue;}
     if((strcmp(argv[i], "-f") == 0) || (strcmp(argv[i], "--data_file") == 0)) {
       if(in.datafile == NULL) in.datafile = new char[1000];

       strcpy(in.datafile, argv[++i]);
       continue;
     }
     if((strcmp(argv[i], "-u") == 0) || (strcmp(argv[i], "--units") == 0)) {
       in.units = strcmp(argv[++i], "metal") == 0 ? 1 : 0;
       continue;
     }

     if((strcmp(argv[i], "-p") == 0) || (strcmp(argv[i], "--force") == 0)) {
       in.forcetype = strcmp(argv[++i], "eam") == 0 ? FORCEEAM : FORCELJ;
       continue;
     }
     if((strcmp(argv[i], "-gn") == 0) || (strcmp(argv[i], "--ghost_newton") == 0)) {
       ghost_newton = atoi(argv[++i]);
       continue;
     }
     if((strcmp(argv[i], "--skip_gpu") == 0)) {
       skip_gpu = atoi(argv[++i]);
       continue;
     }
//.........这里部分代码省略.........

void Integrate::run(Atom &atom, Force* force, Neighbor &neighbor,
                    Comm &comm, Thermo &thermo, Timer &timer)
{
  int i, n;

  comm.timer = &timer;
  timer.array[TIME_TEST] = 0.0;

  int check_safeexchange = comm.check_safeexchange;

  mass = atom.mass;
  dtforce = dtforce / mass;
  //Use OpenMP threads only within the following loop containing the main loop.
  //Do not use OpenMP for setup and postprocessing.
  #pragma omp parallel private(i,n)
  {

    for(n = 0; n < ntimes; n++) {

      #pragma omp barrier

      x = &atom.x[0][0];
      v = &atom.v[0][0];
      f = &atom.f[0][0];
      xold = &atom.xold[0][0];
      nlocal = atom.nlocal;

      initialIntegrate();
      #pragma omp barrier

      #pragma omp master
      timer.stamp();

      if((n + 1) % neighbor.every) {

        #pragma omp barrier
        comm.communicate(atom);
        #pragma omp master
        timer.stamp(TIME_COMM);
        #pragma omp barrier

      } else {
        //these routines are not yet ported to OpenMP
        {
          if(check_safeexchange) {
            #pragma omp master
            {
              double d_max = 0;

              for(i = 0; i < atom.nlocal; i++) {
                double dx = (x[3 * i + 0] - xold[3 * i + 0]);

                if(dx > atom.box.xprd) dx -= atom.box.xprd;

                if(dx < -atom.box.xprd) dx += atom.box.xprd;

                double dy = (x[3 * i + 1] - xold[3 * i + 1]);

                if(dy > atom.box.yprd) dy -= atom.box.yprd;

                if(dy < -atom.box.yprd) dy += atom.box.yprd;

                double dz = (x[3 * i + 2] - xold[3 * i + 2]);

                if(dz > atom.box.zprd) dz -= atom.box.zprd;

                if(dz < -atom.box.zprd) dz += atom.box.zprd;

                double d = dx * dx + dy * dy + dz * dz;

                if(d > d_max) d_max = d;
              }

              d_max = sqrt(d_max);

              if((d_max > atom.box.xhi - atom.box.xlo) || (d_max > atom.box.yhi - atom.box.ylo) || (d_max > atom.box.zhi - atom.box.zlo))
                printf("Warning: Atoms move further than your subdomain size, which will eventually cause lost atoms.\n"
                "Increase reneighboring frequency or choose a different processor grid\n"
                "Maximum move distance: %lf; Subdomain dimensions: %lf %lf %lf\n",
                d_max, atom.box.xhi - atom.box.xlo, atom.box.yhi - atom.box.ylo, atom.box.zhi - atom.box.zlo);

            }

          }


          //int tid = omp_get_thread_num();
          //printf("Check B: %i %i %i\n",comm.me,tid,n);
          #pragma omp master
          timer.stamp_extra_start();
          comm.exchange(atom);
          comm.borders(atom);
          #pragma omp master
          {
            timer.stamp_extra_stop(TIME_TEST);
            timer.stamp(TIME_COMM);
          }

          if(check_safeexchange)
            for(int i = 0; i < 3 * atom.nlocal; i++) atom.xold[i] = atom.x[i];
//.........这里部分代码省略.........

int main(int argc, char** argv)
{
  In in;
  in.datafile = NULL;
  int me = 0;                   //local MPI rank
  int nprocs = 1;               //number of MPI ranks
  int num_threads = 1;		//number of OpenMP threads
  int num_steps = -1;           //number of timesteps (if -1 use value from lj.in)
  int system_size = -1;         //size of the system (if -1 use value from lj.in)
  int nx = -1;
  int ny = -1;
  int nz = -1;
  int check_safeexchange = 0;   //if 1 complain if atom moves further than 1 subdomain length between exchanges
  int do_safeexchange = 0;      //if 1 use safe exchange mode [allows exchange over multiple subdomains]
  int use_sse = 0;              //setting for SSE variant of miniMD only
  int screen_yaml = 0;          //print yaml output to screen also
  int yaml_output = 0;          //print yaml output
  int halfneigh = 1;            //1: use half neighborlist; 0: use full neighborlist; -1: use original miniMD version half neighborlist force
  int teams = 1;
  int device = 0;
  int neighbor_size = -1;
  char* input_file = NULL;
  int ghost_newton = 1;
  int sort = -1;

  for(int i = 0; i < argc; i++) {
    if((strcmp(argv[i], "-i") == 0) || (strcmp(argv[i], "--input_file") == 0)) {
      input_file = argv[++i];
      continue;
    }
  }

  MPI_Init(&argc, &argv);
  MPI_Comm_rank(MPI_COMM_WORLD, &me);
  MPI_Comm_size(MPI_COMM_WORLD, &nprocs);

  int error = 0;

  if(input_file == NULL)
    error = input(in, "in.lj.miniMD");
  else
    error = input(in, input_file);

  if(error) {
    MPI_Finalize();
    exit(0);
  }

  for(int i = 0; i < argc; i++) {
    if((strcmp(argv[i], "-t") == 0) || (strcmp(argv[i], "--num_threads") == 0)) {
      num_threads = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "--teams") == 0)) {
      teams = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "-n") == 0) || (strcmp(argv[i], "--nsteps") == 0))  {
      num_steps = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "-s") == 0) || (strcmp(argv[i], "--size") == 0)) {
      system_size = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "-nx") == 0)) {
      nx = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "-ny") == 0)) {
      ny = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "-nz") == 0)) {
      nz = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "-b") == 0) || (strcmp(argv[i], "--neigh_bins") == 0))  {
      neighbor_size = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "--half_neigh") == 0))  {
      halfneigh = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "-sse") == 0))  {
      use_sse = atoi(argv[++i]);
      continue;
    }

    if((strcmp(argv[i], "--check_exchange") == 0))  {
//.........这里部分代码省略.........