Chris@19: Chris@19:
Chris@19:Chris@19: Previous: Advanced distributed-transpose interface, Chris@19: Up: FFTW MPI Transposes Chris@19:
We close this section by noting that FFTW's MPI transpose routines can
Chris@19: be thought of as a generalization for the MPI_Alltoall
function
Chris@19: (albeit only for floating-point types), and in some circumstances can
Chris@19: function as an improved replacement.
Chris@19:
Chris@19:
Chris@19:
MPI_Alltoall
is defined by the MPI standard as:
Chris@19:
Chris@19:
int MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype, Chris@19: void *recvbuf, int recvcnt, MPI_Datatype recvtype, Chris@19: MPI_Comm comm); Chris@19:Chris@19:
In particular, for double*
arrays in
and out
,
Chris@19: consider the call:
Chris@19:
Chris@19:
MPI_Alltoall(in, howmany, MPI_DOUBLE, out, howmany MPI_DOUBLE, comm); Chris@19:Chris@19:
This is completely equivalent to: Chris@19: Chris@19:
MPI_Comm_size(comm, &P); Chris@19: plan = fftw_mpi_plan_many_transpose(P, P, howmany, 1, 1, in, out, comm, FFTW_ESTIMATE); Chris@19: fftw_execute(plan); Chris@19: fftw_destroy_plan(plan); Chris@19:Chris@19:
That is, computing a P × P transpose on P
processes,
Chris@19: with a block size of 1, is just a standard all-to-all communication.
Chris@19:
Chris@19:
However, using the FFTW routine instead of MPI_Alltoall
may
Chris@19: have certain advantages. First of all, FFTW's routine can operate
Chris@19: in-place (in == out
) whereas MPI_Alltoall
can only
Chris@19: operate out-of-place.
Chris@19:
Chris@19:
Chris@19:
Second, even for out-of-place plans, FFTW's routine may be faster,
Chris@19: especially if you need to perform the all-to-all communication many
Chris@19: times and can afford to use FFTW_MEASURE
or
Chris@19: FFTW_PATIENT
. It should certainly be no slower, not including
Chris@19: the time to create the plan, since one of the possible algorithms that
Chris@19: FFTW uses for an out-of-place transpose is simply to call
Chris@19: MPI_Alltoall
. However, FFTW also considers several other
Chris@19: possible algorithms that, depending on your MPI implementation and
Chris@19: your hardware, may be faster.
Chris@19:
Chris@19:
Chris@19:
Chris@19:
Chris@19: