Mesh_3: bug in perturber

[lrineau]

While debugging something else, a ctest run showed strange failures in another test...

In master (currently 2249dfde063b2d359a93f4e46c9e5ca06789c5f3), if I add the following patch:

diff --git a/Mesh_3/include/CGAL/Mesh_3/C3T3_helpers.h b/Mesh_3/include/CGAL/Mesh_3/C3T3_helpers.h
index 6b3f73f..2ef57f8 100644
--- a/Mesh_3/include/CGAL/Mesh_3/C3T3_helpers.h
+++ b/Mesh_3/include/CGAL/Mesh_3/C3T3_helpers.h
@@ -3374,6 +3374,8 @@ get_least_square_surface_plane(const Vertex_handle& v,

   reference_point = surface_point_vector.front();

+  CGAL_assertion(is_valid(plane.d()));
+
   return plane;
 }

diff --git a/Mesh_3/test/Mesh_3/debug.h b/Mesh_3/test/Mesh_3/debug.h
index 886fa14..30172eb 100644
--- a/Mesh_3/test/Mesh_3/debug.h
+++ b/Mesh_3/test/Mesh_3/debug.h
@@ -17,7 +17,7 @@
 // #define CGAL_MESHES_DEBUG_REFINEMENT_POINTS
 // #define CGAL_DEBUG_OCTREE_CONSTRUCTION

-//#define CGAL_MESH_3_VERBOSE
+#define CGAL_MESH_3_VERBOSE
 //#define CGAL_MESH_3_IO_VERBOSE

 // #define OPTIMIZE_INTERSECTION

then the test test/Mesh_3/test_mesh_implicit_domains.cpp fails. That shows that, for some reason, there are NaN numbers computed in the get_least_square_surface_plane function. I found out that there are "surface centers" with NaN numbers too. Probably some duals are degenerate in floating points, and should be computed robustly.

I failed to reproduce the bug with CGAL-4.9. I do not know why.

[lrineau]

Update: the failing test file is test_mesh_implicit_domains.cpp.

[janetournois]

I cannot reproduce the bug, even though I disabled the time limits on perturb_mesh_3 and exude_mesh_3...

[lrineau]

I reassign this issue to 4.11.

Once we have time, we should try and dig further in the code of the perturber, to find out why NaNs occur the computation.

[lrineau]

Update: Here is the new patch:

diff --git a/Mesh_3/include/CGAL/Mesh_3/C3T3_helpers.h b/Mesh_3/include/CGAL/Mesh_3/C3T3_helpers.h
index 6b3f73f..41e4c06 100644
--- a/Mesh_3/include/CGAL/Mesh_3/C3T3_helpers.h
+++ b/Mesh_3/include/CGAL/Mesh_3/C3T3_helpers.h
@@ -3354,6 +3354,8 @@ get_least_square_surface_plane(const Vertex_handle& v,
       const int& i = fit->second;

       surface_point_vector.push_back(cell->get_facet_surface_center(i));
+      CGAL_assertion_code(const Point_3& p = surface_point_vector.back();)
+      CGAL_assertion(is_valid(p.x()) && is_valid(p.y()) && is_valid(p.z()));
     }
   }

and TBB and C++11 must be involved, for the bug to show up. I can reproduce it more than 90% of time with Mesh_3/test/Mesh_3/test_meshing_3D_image.cpp.

For some reason, I have never seen that bug with a polyhedral domain, only with an implicit one, or with a 3D image, segmented or not. Maybe the fact that the domain oracle is faster increase the change of a race-condition.

[lrineau]

Cc @cjamin

[lrineau]

Actually... I can see the bug even without TBB and the Parallel_tag:

renoir ~/Git/cgal/Mesh_3/test/Mesh_3 $ ../../../build/test/Mesh_3/test_meshing_3D_image
Mesh generation from a 3D image:
        Seed is 1496935255
        Volume is 1736359.78726467
        Number of cells: 3085
        Number of facets: 1291
        Number of vertices: 806
Refining again...
        Number of cells: 3085
        Number of facets: 1291
        Number of vertices: 806
Exude...
        Number of cells: 2938
        Number of facets: 1291
        Number of vertices: 806
        Quality before optimization: 0.268955 - Quality after optimization: 4.8985
        Volume is 1736359.78726467
Perturb...
terminate called after throwing an instance of 'CGAL::Assertion_exception'
  what():  CGAL ERROR: assertion violation!
Expr: is_valid(p.x()) && is_valid(p.y()) && is_valid(p.z())
File: /home/lrineau/Git/cgal/Mesh_3/test/Mesh_3/../../include/CGAL/Mesh_3/C3T3_helpers.h
Line: 3358

The assertion about coordinates being invalid (NaNs) is triggered before the parallel mode is tested.

[lrineau]

@janetournois in the previous message you can see the seed:

Seed is 1496935255

But it seems the issue is barely dependent on the seed, on my Linux machine. My last run was using the seed:

Seed is 1497517282 and it failed the same way.

[lrineau]

After the patch lrineau/cgal@4ddd3f6, I can no longer reproduce the bug without the compiler optimizers.

Actually, before that patch, I do not know if the optimizers were required.

[janetournois]

that at least explains why it is not reproducible on msvc...

[lrineau]

Jane, as I said in my first message in this issue

I failed to reproduce the bug with CGAL-4.9. I do not know why.

So maybe the bug was introduced between 4.9 and 4.10... Maybe related to weighted points?