FFTW 3.3.5: Basic distributed-transpose interface

In particular, suppose that we have an n0 by n1 array in Chris@42: row-major order, block-distributed across the n0 dimension. To Chris@42: transpose this into an n1 by n0 array block-distributed Chris@42: across the n1 dimension, we would create a plan by calling the Chris@42: following function: Chris@42:

Chris@42:

fftw_plan fftw_mpi_plan_transpose(ptrdiff_t n0, ptrdiff_t n1,
Chris@42:                                   double *in, double *out,
Chris@42:                                   MPI_Comm comm, unsigned flags);
Chris@42:

The input and output arrays (in and out) can be the Chris@42: same. The transpose is actually executed by calling Chris@42: fftw_execute on the plan, as usual. Chris@42: Chris@42:

The flags are the usual FFTW planner flags, but support Chris@42: two additional flags: FFTW_MPI_TRANSPOSED_OUT and/or Chris@42: FFTW_MPI_TRANSPOSED_IN. What these flags indicate, for Chris@42: transpose plans, is that the output and/or input, respectively, are Chris@42: locally transposed. That is, on each process input data is Chris@42: normally stored as a local_n0 by n1 array in row-major Chris@42: order, but for an FFTW_MPI_TRANSPOSED_IN plan the input data is Chris@42: stored as n1 by local_n0 in row-major order. Similarly, Chris@42: FFTW_MPI_TRANSPOSED_OUT means that the output is n0 by Chris@42: local_n1 instead of local_n1 by n0. Chris@42: Chris@42: Chris@42:

To determine the local size of the array on each process before and Chris@42: after the transpose, as well as the amount of storage that must be Chris@42: allocated, one should call fftw_mpi_local_size_2d_transposed, Chris@42: just as for a 2d DFT as described in the previous section: Chris@42: Chris@42:

Chris@42:

ptrdiff_t fftw_mpi_local_size_2d_transposed
Chris@42:                 (ptrdiff_t n0, ptrdiff_t n1, MPI_Comm comm,
Chris@42:                  ptrdiff_t *local_n0, ptrdiff_t *local_0_start,
Chris@42:                  ptrdiff_t *local_n1, ptrdiff_t *local_1_start);
Chris@42:

Again, the return value is the local storage to allocate, which in Chris@42: this case is the number of real (double) values rather Chris@42: than complex numbers as in the previous examples. Chris@42:

6.7.1 Basic distributed-transpose interface