[vtk-developers] A new cell type for polyhedral meshes

[Brian Dotson]

Mathieu,
 
OK.  I did a simple test program that brings in an vtkXMLUnstructured
Grid and does a surface filter and then renders and quits.  I run this
test code on data that has been meshed with tets and another that has
been meshed with polyhedrons.  I emphasize, this is the same data that
has been meshed two different ways.  When I run the tetrahedron data
through my program it executes in 4 seconds, when I execute the
polyhedron mesh it takes 4 minutes and 26 seconds.  I ran them both on
callgrind and in the polyhedron case it is spending a lot of time in
LUFactorLinearSystem and LUSolveLinearSystem.
 
I put the code, data, and callgrind output on my ftp site.
 
https://lars.netl.doe.gov/public/dotson/
 
Let me know what you think.
 
Thanks,
 
Brian Dotson
National Energy Technology Laboratory

>>> "Mathieu Malaterre" <mathieu.malaterre at kitware.com> 4/13/2006 5:26
PM >>>

Hi Brian,

    Let's try to do things in order. Can you exhibit a case where the
use 
of vtkConvexPointCell is really slow in a small VTK/C++ code. Ideally 
you should be able to run valgrind+callgrind to analyze what can/should

be optimized.
    A much more complex solution would be to use the Generic framework,
I 
don't know if you are familiar with it, but you can find it in 
VTK/GenericFiltering. Unfortunately it is not as well integrated in 
ParaView. But this is one of our goal.

Anyway let me know how it goes, and what is your time frame ?
Thanks
Mathieu

Brian Dotson wrote:
> Hi all,
> 
>  
> 
> I am working with a group of CFD researchers that are using ParaView
to 
> view their data.  Last year we investigated how ParaView would handle

> polyhedral meshes.  I took their data and created a VTK file that 
> consisted of an unstructured grid with convex point sets for cells. 

> This worked, but when the cell counts and complexities got higher the

> performance dropped considerably.
> 
>  
> 
> I started looking at this problem again, and was wondering if a new
VTK 
> cell type would help this problem.
> 
>  
> 
> My reasoning is that when I am creating the convex point set cells, I
am 
> throwing a lot of useful information away.  When I receive the 
> information from the CFD simulation I know the following:
> 
>  
> 
> 1.  The number of points in each face of the cell.
> 
> 2. What points make up each face of the cell.
> 
> 3.  The points that make up the face are in a specific order.
> 
> 4.  All of the points are on the surface, no internal points.
> 
> 5.  The cell number on each side of the face.
> 
> 6.  The number of faces per cell.
> 
>  
> 
> I was thinking a cell that consisted of a set of faces might be more

> efficient in our application.  Each face would consist of a list of 
> point in a specified order where the normal vector would point away
from 
> the cell. 
> 
>  
> 
> I am willing to do the work on this, but does this make sense?
> 
>  
> 
> Please give me your thoughts.
> 
>  
> 
> Thanks,
> 
>  
> 
> Brian Dotson
> 
> National Energy Technology Laboratory
> 
>  
> 
> 
>
------------------------------------------------------------------------
> 
> _______________________________________________
> vtk-developers mailing list
> vtk-developers at vtk.org
> http://www.vtk.org/mailman/listinfo/vtk-developers

_______________________________________________
vtk-developers mailing list
vtk-developers at vtk.org
http://www.vtk.org/mailman/listinfo/vtk-developers


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/vtk-developers/attachments/20060425/9c73953a/attachment.html>