On 03/09/2015 10:54, Chris Richardson wrote:
This is a mailing list continuation of a bitbucket discussion: https://bitbucket.org/fenics-project/dolfin/pull-requests/240/interpolation-improvementsI have been trying to clean up Function::interpolate() a bit. The following issues have come up: 1. The chain of function calls is long and complex: e.g. to interpolate in the general case from another Function v: - this is just an illustrative sketch, with details removed: Function::interpolate(GenericFunction& v) FunctionSpace::interpolate(Vector& coeffs, const GenericFunction& v) Iterate over cells { v.restrict(cell_coeffs, element, cell, coords, ufc_cell) Function::restrict(w, element, cell, coords, ufc_cell) ...restrict_as_ufc_function(w, element, dolfin_cell, coords, ufc_cell)element.evaluate_dofs() ufc_element->evaluate_dofs() ufc_function::evaluate(vals, coords, ufc_cell) GenericFunction::evaluate(vals, coords, ufc_cell) GenericFunction::eval(vals, x, ufc_cell) GenericFunction::eval(vals, x) Function::eval(vals, x) { id = mesh.bb_tree.compute_first_entity_collision(point) Cell (mesh, id) Function::eval(val, x, cell, ufc_cell) restrict(coeffs, element, cell, coords, ufc_cell) element.evaluate_basis() values += coeffs * basis } 2. Many calls use ufc_cell and dolfin_cell with the same data. This looks bad, and is bad. Can we find a way to transition away from needing both? 3. In parallel, this model totally breaks down, as it is only at the very bottom that we realise the point is 'off process' and we need a collective operation to get it.
4. I forgot to mention, dolfin::Extrapolation::extrapolate() also needs to work in parallel. The name is a bit confusing, and I think it needs better documentation. As I understand it, it is really interpolation with a broader basis, i.e. using surrounding cells. There can be confusion with 'allow_extrapolation' as applied to interpolation, which extrapolates outside the domain. Does that make sense?
Chris _______________________________________________ fenics mailing list fenics@fenicsproject.org http://fenicsproject.org/mailman/listinfo/fenics
