Changeset [35a05dc8f52e4d99d509a7e0171bed07b3b3ee73] by Jed Brown

Implement Q1 sparse matrix assembly for projection problem.

• There is some code duplication computing geometry for sub-elements. Implement this in a generic manner is probably hard, but it should be easy to do better than I did.

• Fix scaling issue where Jdet was not being weighted.

• The convergence rates are somewhat disappointing. While they are nearly mesh independent, a high-tolerance projection is still taking 70 iterations when preconditioned with LU on the sparse mass matrix. I think the Laplacian should actually be better in this regard, due to large gradients being much more localized near the LGL node are large function values. I have not fully ruled out the possibility that there is a mistake, but the Q1 projection matrix fits with by-hand computation for the simple cases and the tensor evaluation demonstrates spectral convergence. They do agree for the Q1 (bdeg=2) case with arbitrary tensor rule.

Signed-off-by: Jed Brown jed@59A2.org http://github.com/jedbrown/dohp/...