Poster: Automatic sparsity determination (ASD)
by
Ralf Giering
Jet-Propuslion-Laboratory
Coauthors: Thomas Kaminski (Max-Plack-Institut fuer Meteorologie, Hamburg)

A new transformation is presented which generates code to compute the sparsity pattern by operations on bit-vectors. This transformation has a forward and reverse mode similar to AD. However, the reverse mode of ASD requires much less intermediate results than its AD counterpart and the forward mode is most efficiently performed in pure mode, which does not evaluate the underlying function itself. ASD is about two orders of magnitude faster than AD. It is applied as a pre-step to exploit the sparsity pattern for optimal seed matricies by graph bi-coloring technique. The final step computes the non-zero Jacobian entries in reverse and forward mode of AD using the two seed matricies.

http://puddle.mit.edu/~ralf/tamc

Date received: December 30, 1999