d@0: d@0: d@0: Previous: Advanced distributed-transpose interface, d@0: Up: FFTW MPI Transposes d@0:
We close this section by noting that FFTW's MPI transpose routines can
d@0: be thought of as a generalization for the MPI_Alltoall function
d@0: (albeit only for floating-point types), and in some circumstances can
d@0: function as an improved replacement.
d@0:
d@0: MPI_Alltoall is defined by the MPI standard as:
d@0:
d@0:
int MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype, d@0: void *recvbuf, int recvcnt, MPI_Datatype recvtype, d@0: MPI_Comm comm); d@0:d@0:
In particular, for double* arrays in and out,
d@0: consider the call:
d@0:
d@0:
MPI_Alltoall(in, howmany, MPI_DOUBLE, out, howmany MPI_DOUBLE, comm); d@0:d@0:
This is completely equivalent to: d@0: d@0:
MPI_Comm_size(comm, &P); d@0: plan = fftw_mpi_plan_many_transpose(P, P, howmany, 1, 1, in, out, comm, FFTW_ESTIMATE); d@0: fftw_execute(plan); d@0: fftw_destroy_plan(plan); d@0:d@0:
That is, computing a P × P transpose on P processes,
d@0: with a block size of 1, is just a standard all-to-all communication.
d@0:
d@0:
However, using the FFTW routine instead of MPI_Alltoall may
d@0: have certain advantages. First of all, FFTW's routine can operate
d@0: in-place in == out) whereas MPI_Alltoall can only
d@0: operate out-of-place.
d@0:
d@0: Second, even for out-of-place plans, FFTW's routine may be faster,
d@0: especially if you need to perform the all-to-all communication many
d@0: times and can afford to use FFTW_MEASURE or
d@0: FFTW_PATIENT. It should certainly be no slower, not including
d@0: the time to create the plan, since one of the possible algorithms that
d@0: FFTW uses for an out-of-place transpose is simply to call
d@0: MPI_Alltoall. However, FFTW also considers several other
d@0: possible algorithms that, depending on your MPI implementation and
d@0: your hardware, may be faster.
d@0:
d@0:
d@0:
d@0:
d@0: