protected virtual void DistributedTranspose(CustomFftPlan data, UnsafeArray src, UnsafeArray dst)
{
FftwMpi.ExecuteR2r(data.TransposePlan.Handler, new IntPtr(src.Ptr), new IntPtr(dst.Ptr));
//int size = data.NumberOfLocalFftsAlongSecondDimension * data.LocalNx;
//_mpi.AllToAll(data.Output.Ptr, size, data.Input.Ptr, size);
//All2All
}
public CustomDistributedFft(Mpi mpi, IProfiler profiler = null)
{
_mpi = mpi;
_profiler = profiler;
Fftw.InitThreads();
FftwMpi.Init();
Fftw.PlanWithNthreads(MultiThreadUtils.MaxDegreeOfParallelism);
}
public FftwPlan CreatePlan(IntPtr memory, int fullNx, int fullNy, int nz, Direction direction, Flags flags)
{
var n = new[] { new IntPtr(fullNx), new IntPtr(fullNy), };
var plan = FftwMpi.PlanManyDft(2, n, new IntPtr(nz), FFTW_MPI_DEFAULT_BLOCK, FFTW_MPI_DEFAULT_BLOCK,
memory, memory, Mpi.CommWorld, (int)direction, (uint)flags);
return(new FftwPlan(fullNx, fullNy, nz, plan));
}
public int GetLocalN0(int fullNx, int fullNy, int nz)
{
var n = new[] { new IntPtr(fullNx), new IntPtr(fullNy), };
IntPtr localN0;
IntPtr localN0Start;
long localSize = FftwMpi.LocalSizeMany(2, n, new IntPtr(nz),
FFTW_MPI_DEFAULT_BLOCK, Mpi.CommWorld, out localN0, out localN0Start).ToInt64();
return(localN0.ToInt32());
}
private void CreateTransposePlan(CustomFftPlan data, uint flags)
{
var n = new IntPtr(2 * data.NumberOfLocalFftsAlongSecondDimension * data.LocalNx);
var np = new IntPtr(data.NumberOfMpiProcesses);
var one = new IntPtr(1);
var input = new IntPtr(data.Input.Ptr);
var output = new IntPtr(data.Output.Ptr);
var handler = FftwMpi.PlanManyTranspose(np, np, n, one, one, input, output, _mpi.Communicator, flags);
data.TransposePlan = new FftwPlan(np.ToInt32(), np.ToInt32(), n.ToInt32(), handler);
}
public void Dispose()
{
FftwMpi.Cleanup();
}
public void ExecutePlan(FftwPlan plan, IntPtr src, IntPtr dst) => FftwMpi.ExecuteDft(plan.Handler, src, dst);