3D mesh generation: Replace operator()(K::Point_3 p) by something with many points?

[nschloe]

Functions like Mesh_domain::create_implicit_mesh_domain are called with domain objects that a have

K::FT
operator()(K::Point_3 p) const
{
  // return the distance to the boundary
}

This is an evaluation returning the signed distance of a point p to the domain boundary.
For mesh generation, this operator is then called with point after point.

For many, if not all of my domains, I could compute the distance much faster if I wasn't forced to do it one after the other, but with many points at once. (Using higher-level BLAS routines, for example.)

Is it possible to make the the mesh generation ask for many points at once?

[MaelRL]

The algorithm works by iteratively inserting points whose position is chosen to best "break" a given element into smaller elements. That position is usually the circumcenter of the Delaunay ball. Thus, this new position is unfortunately dependent on the positions at the step 'n-1'.

However, there could be (at least) two ways to not compute all positions one by one:

• The algorithm has multiple phases: initialization, refinement, optimization. The initial sampling of the domain is computed by shooting random rays and is meant to obtain a triangulation before the circumcenter-based refinement process kicks in. If you know you're going to have many vertices in your mesh, you could greatly increase the amount of points used in the initial sampling, and compute all these points in a batch. Of course, this initial random triangulation will likely be far from honoring your constraint and some refinement (and optimization) will still need be needed.

• At a given step of the refinement, it is likely that all simplices of the mesh must be refined, and thus it is useful to know their circumcenter, which will be the position of some future vertex being inserted. Thus, it might be possible to only compute these circumcenters - and cache them - in a batch once in a while, and when it is required by the refinement algorithm. Unfortunately, that would be quite a few very low level changes...

@lrineau and @janetournois might have better ideas.

[nschloe]

computed by shooting random rays

Since the rays are all independent, they could be computed at the same time.

it is likely that all simplices of the mesh must be refined,

Yes, I assumed that there are operations which take place for all cells individually.

Unfortunately, that would be quite a few very low level changes...

I suppose. In the end, it should come down to looking for loops like

for (auto cell in cells) {
   // evaluate domain.operator()(...)
}

[nschloe]

This came back to bite me again.

For me to understand, do you build the mesh as follows? Build a basic mesh and then refine iteratively, i.e.,

• picking a bad cell,
• inserting a node at the circumcenter (or some other point) of the cell,
• call domain.operator() to keep the new point inside,
• doing some other operations to keep the mesh conforming,
• and repeat.

Do you optimize after each step? Where are the most operator() calls spent?

[janetournois]

For me to understand, do you build the mesh as follows? Build a basic mesh and then refine iteratively, i.e.,

* picking a bad cell,

* inserting a node at the circumcenter (or some other point) of the cell,

the new node does not have to be in the cell, but in its Delaunay ball, to make sure the cell disappears from the mesh

* call `domain.operator()` to keep the new point inside,

This kind of operator is called on newly created cells to keep the "restricted Delaunay triangulation" valid

* doing some other operations to keep the mesh conforming,

the point is inserted in the regular triangulation while keeping it valid, and the "domain status" is updated above. There are no extra-operations here

* and repeat.

Do you optimize after each step? Where are the most operator() calls spent?

Optimization is used only after the end of mesh refinement.

The operator() you mention (I guess you mean the one that updates the cells status wrt the restricted Delaunay triangulation) is called a lot during optimization, because each cell is updated after each move. In global optimizers though, it is possible to compute all moves, apply them all, and then update the full restricted Delaunay triangulation at a time.

[nschloe]

Thanks @janetournois for the reply.

The operator() you mention

This is the operator that, evaluated at a geometrical point x, returns the signed distance to the domain boundary. Did you understand it this way?