[Paraview] vtkDistributedDataFilter and ghost cells

Moreland, Kenneth kmorel at sandia.gov
Tue Nov 6 11:00:55 EST 2012


I believe vtkDistributedDataFilter will always return with global point ids (a mapping from local point ids to global point ids), although it might pass them if they already exist.  So dd->GetOutput()->GetPointData()->GetGlobalIds() should return the array that gives this mapping.

Ghost cells are only created on demand, and this is usually done by pipeline convention.  If you have a filter that needs a layer of ghost cells, it should override the RequestUpdateExtent method to increment the value of UPDATE_NUMBER_OF_GHOST_LEVELS from the output information to the input information.  This method would look something like this.

int vtkDistributedDataFilter::RequestUpdateExtent(
  vtkInformation *vtkNotUsed(request),
  vtkInformationVector **inputVector,
  vtkInformationVector *outputVector)
{
  // get the info objects
  vtkInformation *inInfo = inputVector[0]->GetInformationObject(0);
  vtkInformation *outInfo = outputVector->GetInformationObject(0);

  int piece, numPieces, ghostLevels;

  // We require an extra layer of ghost cells from upstream.

  piece = outInfo->Get(vtkStreamingDemandDrivenPipeline::UPDATE_PIECE_NUMBER());
  numPieces =
    outInfo->Get(vtkStreamingDemandDrivenPipeline::UPDATE_NUMBER_OF_PIECES());
  ghostLevels =
    outInfo->Get(vtkStreamingDemandDrivenPipeline::UPDATE_NUMBER_OF_GHOST_LEVELS());

  inInfo->Set(vtkStreamingDemandDrivenPipeline::UPDATE_PIECE_NUMBER(), piece);
  inInfo->Set(vtkStreamingDemandDrivenPipeline::UPDATE_NUMBER_OF_PIECES(),
              numPieces);
  inInfo->Set(vtkStreamingDemandDrivenPipeline::UPDATE_NUMBER_OF_GHOST_LEVELS(),
              ghostLevels);

  return 1;
}

The operation of the RequestData method should also strip off this layer of ghost cells.  It might be possible to request a layer of ghost cells by setting UPDATE_NUMBER_OF_GHOST_LEVELS at the bottom of the pipeline, but I'm not totally sure how to make that work.  It's probably easier (or at least cleaner) to do it from within a filter.

-Ken

From: Andrew Parker <andy.john.parker at googlemail.com<mailto:andy.john.parker at googlemail.com>>
Date: Tuesday, November 6, 2012 8:25 AM
To: "vtkusers at vtk.org<mailto:vtkusers at vtk.org>" <vtkusers at vtk.org<mailto:vtkusers at vtk.org>>, "paraview at paraview.org<mailto:paraview at paraview.org>" <paraview at paraview.org<mailto:paraview at paraview.org>>
Subject: [EXTERNAL] [Paraview] vtkDistributedDataFilter and ghost cells

Hi,

Hope you can help.  I have some code running in parallel, that by other means I have constructed nprocs worth of vtkRectilinearGrids, one per process.  Each of which is a valid nprocs-worth of the whole serial mesh, I've check this and I am happy with that i.e. it's partitioned properly and nothing is missing.  I need the following information to process my data in parallel:

1) I would like the local -> global cell mapping between the local rgrid and the corresponding global single mesh.
2) I would like to know which cells are on processor boundaries for parallel exchange purposes.
3) I would like all the double arrays per processor to be "expanded" by the amount of (1 level of) ghost cells such that I can properly do the computations I want with the ability to exchange only those additional cells given the local to global mapping.

I have tried from the examples to use the following code, which I call on every process, each of which has it's own local rgrid as I said.  I do the following:

 vtkSmartPointer<vtkDistributedDataFilter> dd = vtkSmartPointer<vtkDistributedDataFilter>::New();
 dd->SetInput(rgrid);

 dd->SetController(getVtkController());
 dd->SetBoundaryModeToSplitBoundaryCells();
 //dd->SetBoundaryModeToAssignToOneRegion();
 //dd->SetBoundaryModeToAssignToAllIntersectingRegions();
 dd->UseMinimalMemoryOff();
 dd->Update();
 vtkPieceScalars *ps = vtkPieceScalars::New();
 ps->SetInputConnection(dd->GetOutputPort());
 ps->SetScalarModeToCellData();
 vtkDataSetSurfaceFilter *dss = vtkDataSetSurfaceFilter::New();
 dss->SetInputConnection(ps->GetOutputPort());

The dd object works fine and writing its contents out on each processor gives nprocs worth of meshes, each of which look slightly different to the way I've partitioned them up, but sum to the same serial mesh so I am happy with that working correctly. But I can't for the life of me figure out how to obtain local to global cell mappings, allocate ghost cells, or work out how to exchange data given the above partition info and comms....

Note I have not provided any additional information to "dd" regarding global cells as per the doxy notes so I assume it went away and computed it.  I can't figure out how to extract it however.  I also have no idea how to modify each local processor rgrid to include the ghost cells for that processor.  Finally given that info, I could exchange between processors to write to each local processors ghost cells the corresponding "real" cell data from the neighbouring meshes and continue the code.

Any help really appreciated!

Cheers,
Andy

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.paraview.org/pipermail/paraview/attachments/20121106/b52d0b87/attachment.htm>


More information about the ParaView mailing list