Improve python exposure of determine_point_ownership
#3344
Conversation
I added a test comparing bounding box tree results to determine_point_ownership. Another possibility would be to look at post files and hardcode rank-cell data into a test. |
| def test_po_against_bb_tree(mesh,points,cells): | ||
| cells = np.array([cell for cell in cells if cell < cell_map.size_local]) | ||
| num_points = points.shape[0] | ||
| po = determine_point_ownership(mesh,points,cells) | ||
| tree = bb_tree(mesh, mesh.topology.dim, cells) |
There was a problem hiding this comment.
Run ruff format and ruff check --fix on the Python files to assure appropriate formatting.
|
Sorry, I made a mess with my branches in my fork and closed the pull request by mistake. Is it possible to reopen it pointing to https://github.com/ordinary-slim/dolfinx/tree/main ? |
I think you have to reopen it from your repository. See: https://gist.github.com/Farhan-Haseeb/fd4cbb476de4e2931c322d7bacb75e5f |
I can't see the PR from my repository. I deleted the original branch of the PR and it automatically closed it. The reopen button is greyed out: The link you provided says:
But I can't find this |
|
I cannot reopen. Please open a new PR:) |
|
Ok, I had to recover the deleted branch with the same name then I could reopen, as per this stackoverflow comment. Sorry for the inconvenience! |
* Python wrapper around nanobind exposed function * Extra optional arguments `cells`
|
@ordinary-slim could you run |
| h = dtype(1.0 / num_cells_side) | ||
| if tdim == 2: | ||
| mesh = create_unit_square( | ||
| MPI.COMM_WORLD, num_cells_side, num_cells_side, CellType.quadrilateral, dtype=dtype |
There was a problem hiding this comment.
Can it be. parameterised across the mesh? (triangle, quad, tet, hex)?
There was a problem hiding this comment.
No, as the test is now it needs the elements to coincide with their bounding boxes. I'll commit a version that covers triangles and tet elements.
| A large padding value will increase the run-time of the code by orders | ||
| of magnitude. General advice is to use a padding on the scale of the | ||
| cell size. | ||
|
|
| /// where src_owner is a list of ranks corresponding to the input | ||
| /// points. dest_owner is a list of ranks corresponding to dest_points, | ||
| /// the points that this process owns. dest_cells contains the | ||
| /// @return Point ownership data with fields `src_owner`, `dest_owner`, `dest_points`, `dest_cells`, |
There was a problem hiding this comment.
I think this documentation should be moved out to the struct itself.
| mesh: Mesh, | ||
| points: npt.NDArray[np.floating], | ||
| cells: typing.Optional[npt.NDArray[np.int32]] = None, | ||
| padding: float = 0.0, |
There was a problem hiding this comment.
As this argument is critical to the behaviour of the function it is best to leave it as a non-default. Then user's need to engage with it, rather than not realise it exists.
There was a problem hiding this comment.
I am bit reluctant to do this change since if cells is left as default then the argument order mesh, points, padding, cells feels a bit awkward to me. Moreover bb_tree also has a default value for padding.
There was a problem hiding this comment.
@jorgensd @garth-wells What do you think about this padding argument? My view is that as it's critical to the operation of the bounding box and we cannot provide a sensible default, it's better not to have it hidden as a default parameter.
There was a problem hiding this comment.
I think it is a crucial parameter. We have already rewritten the API for non-matching interpolation since 0.7, so I think it is fine to make it required.
There was a problem hiding this comment.
@ordinary-slim Please make it non-default argument in both bounding box tree and here.
| /// @note A large padding value can increase the runtime of the function by | ||
| /// orders of magnitude, because for non-colliding cells | ||
| /// one has to determine the closest cell among all processes with an | ||
| /// intersecting bounding box, which is an expensive operation to perform. | ||
| template <std::floating_point T> | ||
| PointOwnershipData<T> determine_point_ownership(const mesh::Mesh<T>& mesh, | ||
| std::span<const T> points, | ||
| T padding) | ||
| std::span<const std::int32_t> cells, | ||
| T padding = 0.0) |
There was a problem hiding this comment.
As this argument is critical to the behaviour of the function it is best to leave it as a non-default. Then user's need to engage with it, rather than not realise it exists.
There was a problem hiding this comment.
padding is forwarded to BoundingBoxTree constructor where it is also default.
There was a problem hiding this comment.
@ordinary-slim Please make it non-default argument in both bounding box tree and here.
| /// @note A large padding value can increase the runtime of the function by | ||
| /// orders of magnitude, because for non-colliding cells | ||
| /// one has to determine the closest cell among all processes with an | ||
| /// intersecting bounding box, which is an expensive operation to perform. | ||
| template <std::floating_point T> | ||
| PointOwnershipData<T> determine_point_ownership(const mesh::Mesh<T>& mesh, | ||
| std::span<const T> points, | ||
| T padding) | ||
| std::span<const std::int32_t> cells, |
There was a problem hiding this comment.
This argument could be converted to std::optional - can be done on later patch.
| of a point if the point is on the surface of a cell in the mesh. | ||
|
|
||
| Returns: | ||
| Point ownership data |
There was a problem hiding this comment.
Are the fields of PointOwnershipData documented in the Python wrapped class?
There was a problem hiding this comment.
dolfinx/python/dolfinx/geometry.py
Lines 47 to 61 in 6fda65d
The doc is the same as on the c++ side.
There was a problem hiding this comment.
Looks good, thanks for pointing it out to me.
The C++ documentation for the return argument describing PointOwnershipData looks over-specified given the struct is now properly documented.
Python wrapper around nanobind exposure of
determine_point_ownershipand extra optional argumentscellsso that the search is performed only forcells.