Both.
But I wouldn't call DOLFINs current system naive. It actually does
pretty much the same thing as the form compiler, except for generating
the code. It builds a similar data structure as FFC (but FFC does a
better job) and precomputes the reference tensor.
/Anders
On Tue, Oct 26, 2004 at 01:28:28PM -0500, Robert C.Kirby wrote:
This is why we need profiling information. Was the computation
dominated by assembly because of intepretive overhead (ie the naive
interpreter bogged down in a deeper tree than for poisson) or because
of increased floating-point demands?
Rob
On Oct 26, 2004, at 12:08 PM, Ridgway Scott wrote:
One example is the module elasticity-updated in DOLFIN for which
assembly time dominates solution time (using the old form evaluation
system in DOLFIN). This would be a typical example of a system which
is more complicated than Poisson and where assembly time may be
significant.
This is a good point. There are these complex-model linear problems
where you just form the matrix and solve. Is that what you were
thinking? Since you can vary the elasticity tensor these days with
special materials, one could even imagine optimizing over varying
elasticity tensors, in an extreme case.
Ridg
