On Thu, 25 Jun 2015 18:50:51 +0200 Jan Blechta <[email protected]> wrote:
> On Thu, 25 Jun 2015 17:59:46 +0200 > Marco Morandini <[email protected]> wrote: > > > On 06/23/2015 04:31 PM, Martin Sandve Alnæs wrote: > > > Both assemble and interpolate have loops over all cells in the > > > mesh, so at least for single-mesh situations the cell will be > > > known. > > > > > > Looking at this from another angle: > > > what exactly is the problem with a Mesh owning its related BBTree? > > > Is it only a matter of making updates safe or is there something > > > else? If we get rid of unsafe ways to modify the Mesh (e.g. > > > MeshGeometry::x()), and add safe ways to modify the Mesh that mark > > > the BBTree as invalidated, does that make all issues go away? > > > > > > > To better understand the issue I tried to write some code. > > > > 1) First option: invalidate Mesh:bonding_box_tree whenever the mesh > > coordinates are accessed with a non-const method > > > > https://bitbucket.org/MarcoMorandini/morandini-dolfin/branch/morandini/invalidate-bbox > > > > I don't know if the implementation is correct (i.e. if I'm > > invalidating it when and only when it is needed), but at first > > glance seems a small change, it passes all the python unit tests, > > and it appears to fix the testcase of > > > > https://bitbucket.org/fenics-project/dolfin/issue/89 > > > > modified to run on current master (attached). > > I think you're missing an invalidation for "MeshGeometry& > Mesh::geometry()". Problem is that Python wrappers cannot distinguish > between > > MeshGeometry& Mesh::geometry() > const MeshGeometry& Mesh::geometry() > > Natural solution seems to be: > > 1. moving tree to MeshGeometry class It could help solving this problem but it seems not being a correct abstraction as tree depends on MeshTopology also. Other solution would be 1. keep tree within Mesh class 2. wrap mesh.geometry() and mesh.coordinates() accessors by wrapper objects which will: a. store pointer to mesh b. call mesh->_valid_tree = false whenever their non-const member is called c. keep mesh->_valid_tree untouched whenever const member is called d. do nothing in the case that mesh is already NULL But this is quite ugly and requires wrappers being properly updated whenever non-const accessor is added to MeshGeometry. Jan > 2. ignoring in SWIG everything which with const and non-const version > 3. handling const-correctness by writing __set__ and __get__ methods > properly > > Johan, any comment here? > > Jan > > > > > > > 2) Second option: actually remove Mesh:bonding_box_tree. > > As pointed out by Logg, this is a more complex task: > > > > https://bitbucket.org/MarcoMorandini/morandini-dolfin/branch/morandini/remove-bbox > > > > The work in not finished, and right now it does not pass all the > > python unit tests. > > > > Just like Garth, I would prefer the solution 2), but I'm starting > > to change my mind. > > > > Anyway. > > Feel free to comment and/or drop everything. > > > > Marco > > > > _______________________________________________ > fenics mailing list > [email protected] > http://fenicsproject.org/mailman/listinfo/fenics _______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
