int main (int argc, char ** argv )
{
int i;
MPI_Init (&argc, &argv);
MPI_Comm_dup (MPI_COMM_WORLD, &comm);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &nproc);
comp_comm_init(comm);
if (processArgs(argc, argv)) {
return 1;
}
if (!rank) {
printf ("Setup parameters:\n");
printf (" # of steps (outputs): %d\n", nsteps);
printf (" # of iterations per step: %d\n", niterations);
printf (" # of computation units in each iteration: %d\n", ncomp);
printf (" # of communication units in each iteration: %d\n", ncomm);
printf (" output size per process: %d x %d doubles = %lld bytes\n",
nx, ny, sizeof(double) * nx * (uint64_t) ny);
printf (" output size per step: %lld bytes\n",
nproc * sizeof(double) * nx * (uint64_t) ny);
}
//2D array with 1D decomposition
offs_x = rank * nx;
offs_y = 0;
gnx = nproc * nx;
gny = ny;
data = (double*) malloc (sizeof(double) * nx * (size_t) ny);
timing_alloc(nsteps);
int bufsizeMB = nx*ny*sizeof(double)/1048576 + 1;
output_init(comm, bufsizeMB);
output_define(nx, ny, gnx, gny, offs_x, offs_y);
int it, step, icomp, icomm;
/* Warm up a bit */
if (!rank) printf ("Warm up for 1 steps, %d iterations per step...\n", niterations);
for (step=0; step < 1; step++) {
for (it=0; it < niterations; it++) {
for (icomp=0; icomp < ncomp; icomp++) {
do_calc_unit (data, nx, ny);
}
for (icomm=0; icomm < ncomm; icomm++) {
do_comm_unit (comm);
}
}
}
/* Do the steps with output now */
data_init();
if (!rank) printf ("Start running with I/O and measurements...\n");
double Tcalc_it, Tcomm_it;
double Truntime; //to print total time for the loop below (for overhead calculation)
char filename[256];
MPI_Barrier (comm);
Truntime = MPI_Wtime();
for (step=0; step < nsteps; step++) {
if (!rank) printf ("Start step %d\n", step);
Tcalc[step] = 0;
Tcomm[step] = 0;
for (it=0; it < niterations; it++) {
// spend some time with computation
Tcalc_it = MPI_Wtime();
for (icomp=0; icomp < ncomp; icomp++) {
do_calc_unit (data, nx, ny);
}
Tcalc_it = MPI_Wtime() - Tcalc_it;
Tcalc[step] += Tcalc_it;
// spend some time with communication
Tcomm_it = MPI_Wtime();
for (icomm=0; icomm < ncomm; icomm++) {
do_comm_unit (comm);
}
Tcomm_it = MPI_Wtime() - Tcomm_it;
Tcomm[step] += Tcomm_it;
}
// output per step
snprintf (filename, sizeof(filename), "data%6.6d", step);
if (!rank) printf ("Output to %s\n", filename);
MPI_Barrier (comm);
output_dump(filename, step, data);
}
MPI_Barrier (comm);
Truntime = MPI_Wtime() - Truntime;
if (!rank) printf ("Finalize...\n");
MPI_Barrier (comm);
output_finalize (rank);
timing_report(nsteps, comm);
double Truntime_max;
MPI_Reduce (&Truntime, &Truntime_max, 1, MPI_DOUBLE, MPI_MAX, 0, comm);
if (!rank) printf ("Total runtime of main loop: %9.3f\n", Truntime);
free (data);
timing_free();
MPI_Barrier (comm);
MPI_Finalize ();
return 0;
}
/*
* Dump the setup code read in the various prologue (.pro and .ps)
* files. The hard part is that we don't want to dump too
* many definitions of fonts, to avoid running out of memory on
* too old PS level 1 printers.
* Nevertheless, I still wait for somebody to tell me if this is
* really needed (useful is sure, needed is not)
*/
void
dump_setup (FILE * stream, a2ps_job * job)
{
output_dump (job->status->setup, stream);
}