Saving cell data for Fides

[lahwaacz]

Since 0771024 there are metadata for the Fides Paraview reader, but the result is very different from the VTX reader opening the same dataset/directory:

• The VTK attribute specifies the data arrays as CellData:

  ADIOS2-Examples/source/cpp/gray-scott/simulation/writer.cpp

  Lines 17 to 23 in 86ed110

  	<CellData Scalars="U">
  	<DataArray Name="U" />
  	<DataArray Name="V" />
  	<DataArray Name="TIME">
  	step
  	</DataArray>
  	</CellData>

• Opening the dataset/directory with the VTX reader results in the following extents (for all x,y,z): 0 to 64 (dimension: 65)
• However, the Fides attributes specify the data as "points":

  ADIOS2-Examples/source/cpp/gray-scott/simulation/writer.cpp

  Lines 67 to 70 in 86ed110

  	std::vector<std::string> varList = {"U", "V"};
  	std::vector<std::string> assocList = {"points", "points"};
  	io.DefineAttribute<std::string>("Fides_Variable_List", varList.data(), varList.size());
  	io.DefineAttribute<std::string>("Fides_Variable_Associations", assocList.data(), assocList.size());

• Opening the same dataset/directory with the Fides reader results in different extents than the VTX reader: 0 to 63 (dimension: 64)
• The VTK and Fides attributes specify different spacing (1 vs 0.1):

  • ADIOS2-Examples/source/cpp/gray-scott/simulation/writer.cpp

    Line 14 in 86ed110

    R"(" Origin="0 0 0" Spacing="1 1 1">

  • ADIOS2-Examples/source/cpp/gray-scott/simulation/writer.cpp

    Lines 63 to 64 in 86ed110

    	double spacing[3] = {0.1, 0.1, 0.1};
    	io.DefineAttribute<double>("Fides_Spacing", &spacing[0], 3);

• Moreover, when the example is run in a distributed manner, e.g. with mpirun -np 8, the Fides dataset contains "gaps" between the subdomains:
  

How can I specify CellData properly for the Fides reader? I tried changing Fides_Variable_Associations from points to cells for all variables, but then the reader fails with the following error and nothing is loaded in Paraview:

ERROR: In vtkFidesReader.cxx, line 697
vtkFidesReader (0x5d464a9f2b20): Couldn't find index of time value 0

It seems that PointData are somewhat necessary for proper dataset loading (maybe Fides_Dimension_Variable?) but there are in fact datasets without any PointData in the real world and it should be possible to load them with Fides.

[lahwaacz]

@kmorel Any ideas regarding these issues? 🙏

[kmorel]

The issue appears to be that field data intended to be associated with cells is actually associated with points. So,

 std::vector<std::string> assocList = {"points", "points"}; 

should probably read

 std::vector<std::string> assocList = {"cells", "cells"}; 

There is a problem with that, though. Currently, your schema is set up to define the extent of points to be the dimensions of the U variable.

ADIOS2-Examples/source/cpp/gray-scott/simulation/writer.cpp

Line 65 in 86ed110

io.DefineAttribute<std::string>("Fides_Dimension_Variable", "U");

Thus, if you associate the variables to the cells, the extents of the cells will not match the extents of the fields.

Ideally, you would have a field associated with points that Fides can use to size things. I'm not sure it can size the grid based on a cell field. However, there might be a way to specify the dimension extents directly.

@caitlinross knows much more about this than I. Hopefully, she can direct you to adjusting the grid dimensions.

[lahwaacz]

Thanks for the input @kmorel. Let me ask another question, rephrasing the problem from the other way: if I have a field that is intended to be worked with as PointData, is there a way to avoid the "gaps" between the subdomains when the dataset is visualized in ParaView? This is motivated by a real problem rather than the ADIOS2-Examples simulation.

[kmorel]

To properly represent point data in a mesh, the points need to be replicated across touching boundaries. That is the only way that cells can be formed on each side. This is not the case for the data being created here. Thus, no process has the points across this gap.

A proper way to implement this is to make the variables associated with the cells. I tried to implement this by creating an attribute that has the point dimensions (which are 1 plus the cell dimensions) and using Fides_Dimensions instead of Fides_Dimension_Variable. However, I ran into a problem where Fides got confused about the number of blocks contained in the data.

At this point I'm going to punt over to @caitlinross. She is the expert on this.