void Grid::initialize(MPI_Comm comm, int nx_, int ny_, int nz_) {
nx = nx_;
ny = ny_;
nz = nz_;
fft_plan = rfftw3d_mpi_create_plan(comm, nx, ny, nz, FFTW_REAL_TO_COMPLEX, FFTW_ESTIMATE);
ifft_plan = rfftw3d_mpi_create_plan(comm, nx, ny, nz, FFTW_COMPLEX_TO_REAL, FFTW_ESTIMATE);
rfftwnd_mpi_local_sizes(fft_plan, &nxloc, &ixmin, &nyloc_t, &iymin_t, &local_size);
#else
void Grid::initialize(int nx_, int ny_, int nz_) {
nx = nx_;
ny = ny_;
nz = nz_;
rfftwnd_plan plan = rfftw3d_create_plan(nx, ny, nz, FFTW_REAL_TO_COMPLEX, FFTW_ESTIMATE | FFTW_IN_PLACE);
rfftwnd_plan iplan = rfftw3d_create_plan(nx, ny, nz, FFTW_COMPLEX_TO_REAL, FFTW_ESTIMATE | FFTW_IN_PLACE);
nxloc = nx;
nyloc_t = ny;
ixmin = iymin_t = 0;
local_size = nx * ny * 2*(nz/2+1);
#endif // HAVE_MPI
/* Allocate extra storage so that each process can hold the boundary
* layer from the adjacent process */
assert(local_size == nxloc*ny*2*(nz/2+1));
local_size = (nxloc+1)*ny*2*(nz/2+1);
data = (fftw_real*) malloc(local_size*sizeof(fftw_real));
}
/*! Initialization of the non-periodic PM routines. The plan-files for FFTW
* are created. Finally, the routine to set-up the non-periodic Greens
* function is called.
*/
void pm_init_nonperiodic(void)
{
int i, slab_to_task_local[GRID];
double bytes_tot = 0;
size_t bytes;
/* Set up the FFTW plan files. */
fft_forward_plan = rfftw3d_mpi_create_plan(MPI_COMM_WORLD, GRID, GRID, GRID,
FFTW_REAL_TO_COMPLEX, FFTW_ESTIMATE | FFTW_IN_PLACE);
fft_inverse_plan = rfftw3d_mpi_create_plan(MPI_COMM_WORLD, GRID, GRID, GRID,
FFTW_COMPLEX_TO_REAL, FFTW_ESTIMATE | FFTW_IN_PLACE);
/* Workspace out the ranges on each processor. */
rfftwnd_mpi_local_sizes(fft_forward_plan, &nslab_x, &slabstart_x, &nslab_y, &slabstart_y, &fftsize);
for(i = 0; i < GRID; i++)
slab_to_task_local[i] = 0;
for(i = 0; i < nslab_x; i++)
slab_to_task_local[slabstart_x + i] = ThisTask;
MPI_Allreduce(slab_to_task_local, slab_to_task, GRID, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
slabs_per_task = malloc(NTask * sizeof(int));
MPI_Allgather(&nslab_x, 1, MPI_INT, slabs_per_task, 1, MPI_INT, MPI_COMM_WORLD);
#ifndef PERIODIC
if(ThisTask == 0)
{
for(i = 0; i < NTask; i++)
printf("Task=%d FFT-Slabs=%d\n", i, slabs_per_task[i]);
}
#endif
first_slab_of_task = malloc(NTask * sizeof(int));
MPI_Allgather(&slabstart_x, 1, MPI_INT, first_slab_of_task, 1, MPI_INT, MPI_COMM_WORLD);
meshmin_list = malloc(3 * NTask * sizeof(int));
meshmax_list = malloc(3 * NTask * sizeof(int));
MPI_Allreduce(&fftsize, &maxfftsize, 1, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
/* now allocate memory to hold the FFT fields */
#if !defined(PERIODIC)
if(!(kernel[0] = (fftw_real *) malloc(bytes = fftsize * sizeof(fftw_real))))
{
printf("failed to allocate memory for `FFT-kernel[0]' (%g MB).\n", bytes / (1024.0 * 1024.0));
endrun(1);
}
bytes_tot += bytes;
fft_of_kernel[0] = (fftw_complex *) kernel[0];
#endif
#if defined(PLACEHIGHRESREGION)
if(!(kernel[1] = (fftw_real *) malloc(bytes = fftsize * sizeof(fftw_real))))
{
printf("failed to allocate memory for `FFT-kernel[1]' (%g MB).\n", bytes / (1024.0 * 1024.0));
endrun(1);
}
bytes_tot += bytes;
fft_of_kernel[1] = (fftw_complex *) kernel[1];
#endif
if(ThisTask == 0)
printf("\nAllocated %g MByte for FFT kernel(s).\n\n", bytes_tot / (1024.0 * 1024.0));
}
/*! This routines generates the FFTW-plans to carry out the parallel FFTs
* later on. Some auxiliary variables are also initialized.
*/
void pm_init_periodic(void)
{
int i;
int slab_to_task_local[PMGRID];
All.Asmth[0] = ASMTH * All.BoxSize / PMGRID;
All.Rcut[0] = RCUT * All.Asmth[0];
/* Set up the FFTW plan files. */
#ifndef NOMPI
fft_forward_plan = rfftw3d_mpi_create_plan(MPI_COMM_WORLD, PMGRID, PMGRID, PMGRID,
FFTW_REAL_TO_COMPLEX, FFTW_ESTIMATE | FFTW_IN_PLACE);
fft_inverse_plan = rfftw3d_mpi_create_plan(MPI_COMM_WORLD, PMGRID, PMGRID, PMGRID,
FFTW_COMPLEX_TO_REAL, FFTW_ESTIMATE | FFTW_IN_PLACE);
#else
fft_forward_plan = rfftw3d_create_plan(PMGRID, PMGRID, PMGRID,
FFTW_REAL_TO_COMPLEX, FFTW_ESTIMATE | FFTW_IN_PLACE);
fft_inverse_plan = rfftw3d_create_plan(PMGRID, PMGRID, PMGRID,
FFTW_COMPLEX_TO_REAL, FFTW_ESTIMATE | FFTW_IN_PLACE);
#endif
/* Workspace out the ranges on each processor. */
rfftwnd_mpi_local_sizes(fft_forward_plan, &nslab_x, &slabstart_x, &nslab_y, &slabstart_y, &fftsize);
for(i = 0; i < PMGRID; i++)
slab_to_task_local[i] = 0;
for(i = 0; i < nslab_x; i++)
slab_to_task_local[slabstart_x + i] = ThisTask;
slabs_per_task = malloc(NTask * sizeof(int));
#ifndef NOMPI
MPI_Allreduce(slab_to_task_local, slab_to_task, PMGRID, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
MPI_Allreduce(&nslab_x, &smallest_slab, 1, MPI_INT, MPI_MIN, MPI_COMM_WORLD);
MPI_Allgather(&nslab_x, 1, MPI_INT, slabs_per_task, 1, MPI_INT, MPI_COMM_WORLD);
#else
slab_to_task = slab_to_task_local;
smallest_slab = nslab_x;
slabs_per_task[0] = nslab_x;
#endif
if(ThisTask == 0)
{
for(i = 0; i < NTask; i++)
printf("Task=%d FFT-Slabs=%d\n", i, slabs_per_task[i]);
}
first_slab_of_task = malloc(NTask * sizeof(int));
MPI_Allgather(&slabstart_x, 1, MPI_INT, first_slab_of_task, 1, MPI_INT, MPI_COMM_WORLD);
meshmin_list = malloc(3 * NTask * sizeof(int));
meshmax_list = malloc(3 * NTask * sizeof(int));
to_slab_fac = PMGRID / All.BoxSize;
#ifndef NOMPI
MPI_Allreduce(&fftsize, &maxfftsize, 1, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
#else
maxfftsize = fftsize;
#endif
}
/*! This routines generates the FFTW-plans to carry out the parallel FFTs
* later on. Some auxiliary variables are also initialized.
*/
void pm_init_periodic(void)
{
int i;
int slab_to_task_local[PMGRID];
All.Asmth[0] = ASMTH * All.BoxSize / PMGRID;
All.Rcut[0] = RCUT * All.Asmth[0];
/* Initialize FFTW MPI */
#ifdef FFTW3
fftw_mpi_init();
#endif
#ifdef FFTW3
/* If using FFTW3, don't create plans yet, just figure out the local array sizes */
fftsize_complex = fftw_mpi_local_size_3d_transposed(PMGRID, PMGRID, 0.5*PMGRID2, MPI_COMM_WORLD, &nslab_x, &slabstart_x, &nslab_y, &slabstart_y);
fftsize_real = 2.*fftsize_complex;
fftw_plan_exists = false;
#else
/* Set up the FFTW plan files. */
fft_forward_plan = rfftw3d_mpi_create_plan(MPI_COMM_WORLD, PMGRID, PMGRID, PMGRID,
FFTW_REAL_TO_COMPLEX, FFTW_ESTIMATE | FFTW_IN_PLACE);
fft_inverse_plan = rfftw3d_mpi_create_plan(MPI_COMM_WORLD, PMGRID, PMGRID, PMGRID,
FFTW_COMPLEX_TO_REAL, FFTW_ESTIMATE | FFTW_IN_PLACE);
/* Workspace out the ranges on each processor. */
rfftwnd_mpi_local_sizes(fft_forward_plan, &nslab_x, &slabstart_x, &nslab_y, &slabstart_y, &fftsize);
#endif
for(i = 0; i < PMGRID; i++)
slab_to_task_local[i] = 0;
for(i = 0; i < nslab_x; i++)
slab_to_task_local[slabstart_x + i] = ThisTask;
MPI_Allreduce(slab_to_task_local, slab_to_task, PMGRID, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
MPI_Allreduce(&nslab_x, &smallest_slab, 1, MPI_INT, MPI_MIN, MPI_COMM_WORLD);
slabs_per_task = malloc(NTask * sizeof(int));
MPI_Allgather(&nslab_x, 1, MPI_INT, slabs_per_task, 1, MPI_INT, MPI_COMM_WORLD);
if(ThisTask == 0)
{
for(i = 0; i < NTask; i++)
printf("Task=%d FFT-Slabs=%d\n", i, slabs_per_task[i]);
}
first_slab_of_task = malloc(NTask * sizeof(int));
MPI_Allgather(&slabstart_x, 1, MPI_INT, first_slab_of_task, 1, MPI_INT, MPI_COMM_WORLD);
meshmin_list = malloc(3 * NTask * sizeof(int));
meshmax_list = malloc(3 * NTask * sizeof(int));
to_slab_fac = PMGRID / All.BoxSize;
MPI_Allreduce(&fftsize, &maxfftsize, 1, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
}
