void Allocate()
{
CheckParametres();
EEE = new std::complex<double>[local_nx*NN*NN_t];
sp_coeff = new std::complex<double>[NN_t];
fluence = new double[local_nx*NN];
CreateDispersionCoeff(sp_coeff, NN_t, ldiff, ldisp);
plan_f = fftw3d_mpi_create_plan(MPI_COMM_WORLD, NN, NN, NN_t, FFTW_FORWARD, FFTW_ESTIMATE);
plan_b = fftw3d_mpi_create_plan(MPI_COMM_WORLD, NN, NN, NN_t, FFTW_BACKWARD, FFTW_ESTIMATE);
MPI_Barrier(MPI_COMM_WORLD);
}
void testnd_in_place(int rank, int *n, fftw_direction dir,
fftwnd_plan validated_plan,
int alternate_api, int specific, int force_buffered)
{
int local_nx, local_x_start, local_ny_after_transpose,
local_y_start_after_transpose, total_local_size;
int istride;
int N, dim, i;
fftw_complex *in1, *work = 0, *in2;
fftwnd_mpi_plan p = 0;
int flags = measure_flag | wisdom_flag | FFTW_IN_PLACE;
if (specific || rank < 2)
return;
if (coinflip())
flags |= FFTW_THREADSAFE;
if (force_buffered)
flags |= FFTWND_FORCE_BUFFERED;
N = 1;
for (dim = 0; dim < rank; ++dim)
N *= n[dim];
if (alternate_api && (rank == 2 || rank == 3)) {
if (rank == 2)
p = fftw2d_mpi_create_plan(MPI_COMM_WORLD,
n[0], n[1], dir, flags);
else
p = fftw3d_mpi_create_plan(MPI_COMM_WORLD,
n[0], n[1], n[2], dir, flags);
}
else /* standard api */
p = fftwnd_mpi_create_plan(MPI_COMM_WORLD, rank, n, dir, flags);
fftwnd_mpi_local_sizes(p, &local_nx, &local_x_start,
&local_ny_after_transpose,
&local_y_start_after_transpose,
&total_local_size);
in1 = (fftw_complex *) fftw_malloc(total_local_size * MAX_STRIDE
* sizeof(fftw_complex));
if (coinflip()) {
WHEN_VERBOSE(1, my_printf("w/work..."));
work = (fftw_complex *) fftw_malloc(total_local_size * MAX_STRIDE
* sizeof(fftw_complex));
}
in2 = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
for (istride = 1; istride <= MAX_STRIDE; ++istride) {
/* generate random inputs */
for (i = 0; i < N; ++i) {
c_re(in2[i]) = DRAND();
c_im(in2[i]) = DRAND();
}
for (i = 0; i < local_nx * (N/n[0]); ++i) {
int j;
for (j = 0; j < istride; ++j) {
c_re(in1[i * istride + j]) = c_re((in2 + local_x_start
* (N/n[0])) [i]);
c_im(in1[i * istride + j]) = c_im((in2 + local_x_start
* (N/n[0])) [i]);
}
}
fftwnd_mpi(p, istride, in1, work, FFTW_NORMAL_ORDER);
fftwnd(validated_plan, 1, in2, 1, 1, NULL, 0, 0);
for (i = 0; i < istride; ++i)
CHECK(compute_error_complex(in1 + i, istride,
in2 + local_x_start * (N/n[0]),
1, local_nx * (N/n[0])) < TOLERANCE,
"testnd_in_place: wrong answer");
}
fftwnd_mpi_destroy_plan(p);
fftw_free(in2);
fftw_free(work);
fftw_free(in1);
}
