d@0: d@0: d@0: Next: Advanced distributed-transpose interface, d@0: Previous: FFTW MPI Transposes, d@0: Up: FFTW MPI Transposes d@0:
In particular, suppose that we have an n0 by n1 array in
d@0: row-major order, block-distributed across the n0 dimension. To
d@0: transpose this into an n1 by n0 array block-distributed
d@0: across the n1 dimension, we would create a plan by calling the
d@0: following function:
d@0:
d@0:
fftw_plan fftw_mpi_plan_transpose(ptrdiff_t n0, ptrdiff_t n1, d@0: double *in, double *out, d@0: MPI_Comm comm, unsigned flags); d@0:d@0:
d@0: The input and output arrays (in and out) can be the
d@0: same. The transpose is actually executed by calling
d@0: fftw_execute on the plan, as usual.
d@0:
d@0: The flags are the usual FFTW planner flags, but support
d@0: two additional flags: FFTW_MPI_TRANSPOSED_OUT and/or
d@0: FFTW_MPI_TRANSPOSED_IN. What these flags indicate, for
d@0: transpose plans, is that the output and/or input, respectively, are
d@0: locally transposed. That is, on each process input data is
d@0: normally stored as a local_n0 by n1 array in row-major
d@0: order, but for an FFTW_MPI_TRANSPOSED_IN plan the input data is
d@0: stored as n1 by local_n0 in row-major order. Similarly,
d@0: FFTW_MPI_TRANSPOSED_OUT means that the output is n0 by
d@0: local_n1 instead of local_n1 by n0.
d@0:
d@0: To determine the local size of the array on each process before and
d@0: after the transpose, as well as the amount of storage that must be
d@0: allocated, one should call fftw_mpi_local_size_2d_transposed,
d@0: just as for a 2d DFT as described in the previous section:
d@0:
d@0:
ptrdiff_t fftw_mpi_local_size_2d_transposed d@0: (ptrdiff_t n0, ptrdiff_t n1, MPI_Comm comm, d@0: ptrdiff_t *local_n0, ptrdiff_t *local_0_start, d@0: ptrdiff_t *local_n1, ptrdiff_t *local_1_start); d@0:d@0: