int main(int argc, char **argv)
{
int np[2];
ptrdiff_t n[4];
ptrdiff_t alloc_local;
ptrdiff_t local_ni[4], local_i_start[4];
ptrdiff_t local_no[4], local_o_start[4];
double err, *in;
pfft_complex *out;
pfft_plan plan_forw=NULL, plan_back=NULL;
MPI_Comm comm_cart_2d;
/* Set size of FFT and process mesh */
n[0] = 13; n[1] = 14; n[2] = 19; n[3] = 17;
np[0] = 2; np[1] = 2;
/* Initialize MPI and PFFT */
MPI_Init(&argc, &argv);
pfft_init();
/* Create two-dimensional process grid of size np[0] x np[1], if possible */
if( pfft_create_procmesh_2d(MPI_COMM_WORLD, np[0], np[1], &comm_cart_2d) ){
pfft_fprintf(MPI_COMM_WORLD, stderr, "Error: This test file only works with %d processes.\n", np[0]*np[1]);
MPI_Finalize();
return 1;
}
/* Get parameters of data distribution */
alloc_local = pfft_local_size_dft_r2c(4, n, comm_cart_2d, PFFT_TRANSPOSED_OUT,
local_ni, local_i_start, local_no, local_o_start);
/* Allocate memory */
in = pfft_alloc_real(2 * alloc_local);
out = pfft_alloc_complex(alloc_local);
/* Plan parallel forward FFT */
plan_forw = pfft_plan_dft_r2c(
4, n, in, out, comm_cart_2d, PFFT_FORWARD, PFFT_TRANSPOSED_OUT| PFFT_MEASURE| PFFT_DESTROY_INPUT);
/* Plan parallel backward FFT */
plan_back = pfft_plan_dft_c2r(
4, n, out, in, comm_cart_2d, PFFT_BACKWARD, PFFT_TRANSPOSED_IN| PFFT_MEASURE| PFFT_DESTROY_INPUT);
/* Initialize input with random numbers */
pfft_init_input_real(4, n, local_ni, local_i_start,
in);
/* execute parallel forward FFT */
pfft_execute(plan_forw);
/* execute parallel backward FFT */
pfft_execute(plan_back);
/* Scale data */
for(ptrdiff_t l=0; l < local_ni[0] * local_ni[1] * local_ni[2] * local_ni[3]; l++)
in[l] /= (n[0]*n[1]*n[2]*n[3]);
/* Print error of back transformed data */
MPI_Barrier(MPI_COMM_WORLD);
err = pfft_check_output_real(4, n, local_ni, local_i_start, in, comm_cart_2d);
pfft_printf(comm_cart_2d, "Error after one forward and backward trafo of size n=(%td, %td, %td, %td):\n", n[0], n[1], n[2], n[3]);
pfft_printf(comm_cart_2d, "maxerror = %6.2e;\n", err);
/* free mem and finalize */
pfft_destroy_plan(plan_forw);
pfft_destroy_plan(plan_back);
MPI_Comm_free(&comm_cart_2d);
pfft_free(in); pfft_free(out);
MPI_Finalize();
return 0;
}
void pm_init(PM * pm, PMInit * init, MPI_Comm comm) {
pm->init = *init;
pm->mem = _libfastpm_get_gmem();
/* initialize the domain */
MPI_Comm_rank(comm, &pm->ThisTask);
MPI_Comm_size(comm, &pm->NTask);
int Ny = init->NprocY;
int Nx;
if(Ny <= 0) {
Ny = 1;
Nx = pm->NTask;
if(!init->use_fftw) {
for(; Ny * Ny < pm->NTask; Ny ++) continue;
for(; Ny >= 1; Ny--) {
if (pm->NTask % Ny == 0) break;
continue;
}
}
} else {
if(pm->NTask % Ny != 0) {
fastpm_raise(-1, "NprocY(%d) and NTask(%d) is incompatible\n", Ny, pm->NTask);
}
}
Nx = pm->NTask / Ny;
pm->Nproc[0] = Nx;
pm->Nproc[1] = Ny;
if(init->use_fftw) {
if(Ny != 1) {
fastpm_raise(-1, "FFTW requires Ny == 1; Ny = %d\n", Ny);
}
}
int d;
pm->Norm = 1.0;
pm->Volume = 1.0;
for(d = 0; d < 3; d ++) {
pm->Nmesh[d] = init->Nmesh;
pm->BoxSize[d] = init->BoxSize;
pm->Below[d] = 0;
pm->Above[d] = 1;
pm->CellSize[d] = pm->BoxSize[d] / pm->Nmesh[d];
pm->InvCellSize[d] = 1.0 / pm->CellSize[d];
pm->Norm *= pm->Nmesh[d];
pm->Volume *= pm->BoxSize[d];
}
pfft_create_procmesh(2, comm, pm->Nproc, &pm->Comm2D);
if(init->use_fftw) {
pm->allocsize = 2 * fftw_local_size_dft_r2c(
3, pm->Nmesh, pm->Comm2D,
(pm->init.transposed?FFTW_MPI_TRANSPOSED_OUT:0),
pm->IRegion.size, pm->IRegion.start,
pm->ORegion.size, pm->ORegion.start);
} else {
pm->allocsize = 2 * pfft_local_size_dft_r2c(
3, pm->Nmesh, pm->Comm2D,
(pm->init.transposed?PFFT_TRANSPOSED_OUT:0)
| PFFT_PADDED_R2C,
pm->IRegion.size, pm->IRegion.start,
pm->ORegion.size, pm->ORegion.start);
}
/* Note that we need to fix up the padded size of the real data;
* and transpose with strides , */
pm->IRegion.strides[2] = 1;
pm->IRegion.strides[1] = pm->IRegion.size[2];
pm->IRegion.strides[0] = pm->IRegion.size[1] * pm->IRegion.strides[1];
pm->IRegion.total = pm->IRegion.size[0] * pm->IRegion.strides[0];
/* remove padding from the view */
pm->IRegion.size[2] = pm->Nmesh[2];
if(pm->init.transposed) {
if(pm->init.use_fftw) {
/* FFTW transposed, y, x, z */
pm->ORegion.strides[2] = 1;
pm->ORegion.strides[0] = pm->ORegion.size[2];
pm->ORegion.strides[1] = pm->ORegion.size[0] * pm->ORegion.strides[0];
pm->ORegion.total = pm->ORegion.size[1] * pm->ORegion.strides[1];
} else {
/* PFFT transposed, y, z, x */
pm->ORegion.strides[0] = 1;
pm->ORegion.strides[2] = pm->ORegion.size[0];
pm->ORegion.strides[1] = pm->ORegion.size[2] * pm->ORegion.strides[2];
pm->ORegion.total = pm->ORegion.size[1] * pm->ORegion.strides[1];
}
} else {
/* non-transposed */
pm->ORegion.strides[2] = 1;
pm->ORegion.strides[1] = pm->ORegion.size[2];
pm->ORegion.strides[0] = pm->ORegion.size[1] * pm->ORegion.strides[1];
pm->ORegion.total = pm->ORegion.size[0] * pm->ORegion.strides[0];
}
for(d = 0; d < 2; d ++) {
MPI_Comm projected;
int remain_dims[2] = {0, 0};
remain_dims[d] = 1;
pm->Grid.edges_int[d] =
malloc(sizeof(pm->Grid.edges_int[0][0]) * (pm->Nproc[d] + 1));
pm->Grid.edges_float[d] =
malloc(sizeof(pm->Grid.edges_float[0][0]) * (pm->Nproc[d] + 1));
pm->Grid.MeshtoCart[d] = malloc(sizeof(int) * pm->Nmesh[d]);
MPI_Cart_sub(pm->Comm2D, remain_dims, &projected);
MPI_Allgather(&pm->IRegion.start[d], 1, MPI_PTRDIFF,
pm->Grid.edges_int[d], 1, MPI_PTRDIFF, projected);
int ntask;
MPI_Comm_size(projected, &ntask);
MPI_Comm_free(&projected);
int j;
for(j = 0; j < pm->Nproc[d]; j ++) {
pm->Grid.edges_float[d][j] = 1.0 * pm->Grid.edges_int[d][j] / pm->Nmesh[d] * pm->BoxSize[d];
}
/* Last edge is at the edge of the box */
pm->Grid.edges_float[d][j] = pm->BoxSize[d];
pm->Grid.edges_int[d][j] = pm->Nmesh[d];
/* fill in the look up table */
for(j = 0; j < pm->Nproc[d]; j ++) {
int i;
for(i = pm->Grid.edges_int[d][j]; i < pm->Grid.edges_int[d][j+1]; i ++) {
pm->Grid.MeshtoCart[d][i] = j;
}
}
}
FastPMFloat * canvas = pm_alloc(pm);
FastPMFloat * workspace = pm_alloc(pm);
if(pm->init.use_fftw) {
pm->r2c = plan_dft_r2c_fftw(
3, pm->Nmesh, (void*) workspace, (void*) canvas,
pm->Comm2D,
(pm->init.transposed?FFTW_MPI_TRANSPOSED_OUT:0)
| FFTW_ESTIMATE
| FFTW_DESTROY_INPUT
);
pm->c2r = plan_dft_c2r_fftw(
3, pm->Nmesh, (void*) canvas, (void*) canvas,
pm->Comm2D,
(pm->init.transposed?FFTW_MPI_TRANSPOSED_IN:0)
| FFTW_ESTIMATE
| FFTW_DESTROY_INPUT
);
} else {
pm->r2c = plan_dft_r2c(
3, pm->Nmesh, (void*) workspace, (void*) canvas,
pm->Comm2D,
PFFT_FORWARD,
(pm->init.transposed?PFFT_TRANSPOSED_OUT:0)
| PFFT_PADDED_R2C
| PFFT_ESTIMATE
| PFFT_TUNE
//| PFFT_MEASURE
| PFFT_DESTROY_INPUT
);
pm->c2r = plan_dft_c2r(
3, pm->Nmesh, (void*) workspace, (void*) workspace,
pm->Comm2D,
PFFT_BACKWARD,
(pm->init.transposed?PFFT_TRANSPOSED_IN:0)
| PFFT_PADDED_C2R
| PFFT_ESTIMATE
//| PFFT_MEASURE
| PFFT_TUNE
| PFFT_DESTROY_INPUT
);
}
pm_free(pm, workspace);
pm_free(pm, canvas);
for(d = 0; d < 3; d++) {
pm->MeshtoK[d] = malloc(pm->Nmesh[d] * sizeof(double));
int i;
for(i = 0; i < pm->Nmesh[d]; i++) {
int ii = i;
if(ii >= pm->Nmesh[d] / 2) {
ii -= pm->Nmesh[d];
}
pm->MeshtoK[d][i] = ii * 2 * M_PI / pm->BoxSize[d];
}
}
}
