Replace .at() interpolation in 2D callbacks #373
Conversation
- .at() raises an error when the number of processors used is large, use a VertexOnlyMesh instead for interpolation
|
Prevents the Firedrake issue #2790 whilst Firedrake issue #3080 is waiting to be resolved. This method is slower than There are some remaining direct applications of |
thetis/callback.py
Outdated
| self.vom = VertexOnlyMesh(self.mesh, detector_locations, redundant=True) | ||
| self.function_spaces = {} | ||
| for field_name in field_names: | ||
| field = solver_obj.fields[field_name] | ||
| if isinstance(field.function_space().ufl_element(), VectorElement): | ||
| P0DG = VectorFunctionSpace(self.vom, "DG", 0) | ||
| P0DG_input_ordering = VectorFunctionSpace(self.vom.input_ordering, "DG", 0) | ||
| else: | ||
| P0DG = FunctionSpace(self.vom, "DG", 0) | ||
| P0DG_input_ordering = FunctionSpace(self.vom.input_ordering, "DG", 0) | ||
| self.function_spaces[field_name] = (P0DG, P0DG_input_ordering) |
There was a problem hiding this comment.
minor: there's quite a lot of duplicated code in the vom creation and evaluation. Could plausibly be refactored in to helper functions for example
There was a problem hiding this comment.
I've created a new function in utility.py L1159: vom_interpolator_function. It generates and returns the Functions only as suggested so we remove all the unnecessary Function duplication I had. We could probably go a step further and create a vom_interpolator class which also performs the interpolation as well rather than repeating these lines:
f_at_points, f_at_input_points = self.interp_functions[field_name]
f_at_points.interpolate(field)
f_at_input_points.interpolate(f_at_points)
return f_at_input_points.dat.data_ro[:]
Though maybe it's fine how it is for now. When Firedrake issue 3080 is dealt with we can probably revert to .at() anyway.
| f_at_points.interpolate(field) | ||
| f_at_input_points.interpolate(f_at_points) |
There was a problem hiding this comment.
Do you know if this vom evaluation works for 2D mesh on the sphere (in which case the point coordinates are 3d x,y,z)? The whole eval_func option was added to enable at()-like evaluation in this case (eval_func was a modified version of Firedrake's at)
There was a problem hiding this comment.
Unsure as I only usually do boring plain 2D meshes! I mainly fixed because of the parallelisation issues for big simulations. Happy to look into it in the future or perhaps someone in the Firedrake group might have answer for this already.
|
The current version of Firedrake is failing, which seems to be causing all the Thetis tests to fail. Waiting for this to be updated. |
.at() raises an error when the number of processors used is large, use a VertexOnlyMesh instead for interpolation