updates after 2012.3

[Robert Cimrman]

Hi,

this is to inform you about the latest updates of the code.

1. The 'hierarchic' branch has been merged into the master. This branch
introduces a hierarchical FE basis into the code and some other tools described
below. Look at [1] for an example. It's not finished, but I have decided to
merge it already, as the general ideas seem to be work, and I needed some of
the developments in the master and vice-versa. The caveats are for now:

- Only 2_4 (rectangles) and 3_8 (bricks) elements are supported.
- Dirichlet boundary conditions are piece-wise linear - the higher order
boundary dofs are set to zero. To allow them, projections to FE spaces on a
region would have to be implemented. We are thinking about redesigning the C FE
core to make tasks such as this easier.
- Surface integral terms are not yet supported.
- The post-processing is more difficult, as a linearization of the solution is
required.
- A test was written that checks the basis continuity between two elements.
This test revealed that the standard nodal basis on 3_8 elements has wrong face
orientations (the functions from the two face sides do not match) for
approximation order >= 3. This will be fixed, hopefully before the next release.

2. New plotting modules / updated scripts:

- script/save_basis.py can be used to visualize a FE basis or its gradient on a
single element or on a (small) mesh.
- New modules sfepy/postprocess/plot_facets.py and
sfepy/postprocess/plot_dofs.py can be used to plot the reference elements and
DOF numberings. Matplotlib with mplot3d is needed there.

r.
[1] http://docs.sfepy.org/doc-devel/examples/diffusion/sinbc.html

[Ondřej Čertík]

On Wed, Sep 26, 2012 at 9:20 AM, Robert Cimrman <cimr...@ntc.zcu.cz> wrote:
> Hi,
>
> this is to inform you about the latest updates of the code.
>
> 1. The 'hierarchic' branch has been merged into the master. This branch
> introduces a hierarchical FE basis into the code and some other tools
> described below. Look at [1] for an example. It's not finished, but I have
> decided to merge it already, as the general ideas seem to be work, and I
> needed some of the developments in the master and vice-versa. The caveats
> are for now:
>
> - Only 2_4 (rectangles) and 3_8 (bricks) elements are supported.

That's exciting. Is it using the Cartesian product on each element?
What is the maximum polynomial order that it can handle?

Ondrej

[Robert Cimrman]

Yes, but the order in each direction as well as in all the elements is the same.

> What is the maximum polynomial order that it can handle?

The 1D base functions are generated up to order 10. But there is a limitation
given by the fact that we assemble all elements in a group at once, so
unintegrated element matrices have to fit into memory. In 3D for the 1D order 7
the element matrix has shape 512x512, and the number of quadrature points is
also 512 so 8 * 512**3 / 1e9 = 1.073741824, that is 1GB for a single element.
This will have to be dealt with.

r.

[Ondřej Čertík]

I see. Looks like one cannot precalculate everything on quadrature
points in advance.
So I guess one will have to deal with it element by element.

Ondrej