New decimator #1175

elalish · 2025-02-28T20:57:38Z

My idea here is to speed up our Simplify() function by triangulating the identified faces in parallel rather than collapsing the edges serially. I was also hoping this would produce a better triangulation with fewer slivers, since that makes more difference in a situation like this where the verts of a face are not entirely coplanar. ~~That part doesn't seem to be true so far, but this may point toward ways to improve the triangulator.~~ The problem was the fast convex triangulator, which makes poorer quality triangles.

We went from this in 3 seconds (8 million triangles to 2,500):

To doing the same thing in 1.5 seconds with this result:

~~At the moment this appears to be working in serial mode, but is failing in parallel, so I've probably messed something up with my atomics.~~ Fixed.

elalish · 2025-02-28T20:59:31Z

src/impl.h

@@ -324,9 +324,7 @@ struct Manifold::Impl {

  // face_op.cpp
  void Face2Tri(const Vec<int>& faceEdge, const Vec<TriRef>& halfedgeRef);
-  PolygonsIdx Face2Polygons(VecView<Halfedge>::IterC start,
-                            VecView<Halfedge>::IterC end,
-                            mat2x3 projection) const;


I split Face2Polygons into two non-member functions: AssembleHalfedges and ProjectPolygons.

pca006132

I don't see anything wrong with synchronization here. Probably need more time to debug.

src/face_op.cpp

pca006132 · 2025-03-01T08:44:00Z

For BooleanComplex.CraycloudBool, it seems that the flatten face procedure can get no faces, making the other logic fail. However, I am not entirely sure why it can flatten the whole mesh into nothing...

elalish · 2025-03-01T22:39:31Z

Ah, thanks for noticing that - I'd say it's actually desired behavior. Ideally I'd like any mesh with effectively zero volume (not an easy thing to check in practice, but conceptually) to get decimated down to empty. I think that means any flattish-mesh with a thickness less than 2x tolerance will find only two faces, and no verts (since a vert is an intersection of 3 or more faces). I'll toss in some logic to early-out in that case.

…cimator

elalish · 2025-03-02T05:05:53Z

src/face_op.cpp

+    if (numEdge > 0) {
+      std::copy(polys2.begin(), polys2.end(), newHalfedge2.begin() + start);
+    }
+    faceEdge2[face + 1] = start + numEdge;


I found the parallel problem - can't think of the correct way to do this right now...

Is the order in newHalfedge2 important? Can we fix them later instead?

Oh ok I get it, faceEdge requires following the order. I guess we could just modify the definition a bit...

pca006132 · 2025-03-02T12:44:16Z

Ugly hack to make it work:

diff --git i/src/boolean_result.cpp w/src/boolean_result.cpp
index afc0a6ee..7493b81c 100644
--- i/src/boolean_result.cpp
+++ w/src/boolean_result.cpp
@@ -844,7 +844,7 @@ Manifold::Impl Boolean3::Result(OpType op) const {
   if (ManifoldParams().intermediateChecks)
     DEBUG_ASSERT(outR.IsManifold(), logicErr, "polygon mesh is not manifold!");
 
-  outR.Face2Tri(faceEdge, halfedgeRef);
+  outR.Face2Tri(faceEdge, halfedgeRef, {});
   halfedgeRef.clear();
   faceEdge.clear();
 
diff --git i/src/face_op.cpp w/src/face_op.cpp
index 62c8d026..230a0cd2 100644
--- i/src/face_op.cpp
+++ w/src/face_op.cpp
@@ -91,7 +91,8 @@ using AddTriangle = std::function<void(int, ivec3, vec3, TriRef)>;
  * faceNormal_ values are retained, repeated as necessary.
  */
 void Manifold::Impl::Face2Tri(const Vec<int>& faceEdge,
-                              const Vec<TriRef>& halfedgeRef) {
+                              const Vec<TriRef>& halfedgeRef,
+                              const Vec<int>& faceEdgeStart) {
   ZoneScoped;
   Vec<ivec3> triVerts;
   Vec<vec3> triNormal;
@@ -99,7 +100,7 @@ void Manifold::Impl::Face2Tri(const Vec<int>& faceEdge,
   triRef.clear();
   auto processFace = [&](GeneralTriangulation general, AddTriangle addTri,
                          int face) {
-    const int firstEdge = faceEdge[face];
+    const int firstEdge = faceEdgeStart.empty() ? faceEdge[face] : faceEdgeStart[face];
     const int lastEdge = faceEdge[face + 1];
     const int numEdge = lastEdge - firstEdge;
     if (numEdge == 0) return;
@@ -181,7 +182,7 @@ void Manifold::Impl::Face2Tri(const Vec<int>& faceEdge,
     const vec3 normal = faceNormal_[face];
     const mat2x3 projection = GetAxisAlignedProjection(normal);
     const PolygonsIdx polys = ProjectPolygons(
-        AssembleHalfedges(halfedge_.cbegin() + faceEdge[face],
+        AssembleHalfedges(halfedge_.cbegin() + (faceEdgeStart.empty() ? faceEdge[face] : faceEdgeStart[face]),
                           halfedge_.cbegin() + faceEdge[face + 1]),
         vertPos_, projection);
     return TriangulateIdx(polys, epsilon_);
@@ -196,7 +197,7 @@ void Manifold::Impl::Face2Tri(const Vec<int>& faceEdge,
   // triangulate complex faces
   for_each(autoPolicy(faceEdge.size(), 1e5), countAt(0_uz),
            countAt(faceEdge.size() - 1), [&](size_t face) {
-             triCount[face] = faceEdge[face + 1] - faceEdge[face] - 2;
+             triCount[face] = faceEdge[face + 1] - (faceEdgeStart.empty() ? faceEdge[face] : faceEdgeStart[face]) - 2;
              DEBUG_ASSERT(triCount[face] >= 1, topologyErr,
                           "face has less than three edges.");
              if (triCount[face] > 2)
@@ -307,6 +308,7 @@ void Manifold::Impl::FlattenFaces() {
 
   std::atomic<int> startFace(0);
   Vec<int> faceEdge2(faceEdge.size());
+  Vec<int> faceEdgeStart(faceEdge.size() - 1);
   faceEdge2[0] = 0;
   Vec<Halfedge> newHalfedge2(newHalfedge.size());
   for_each_n(policy, countAt(0_uz), numFace, [&](size_t face) {
@@ -334,6 +336,7 @@ void Manifold::Impl::FlattenFaces() {
     if (numEdge > 0) {
       std::copy(polys2.begin(), polys2.end(), newHalfedge2.begin() + start);
     }
+    faceEdgeStart[face] = start;
     faceEdge2[face + 1] = start + numEdge;
 
     const int oldFace = newHalf2Old[faceEdge[face]] / 3;
@@ -350,7 +353,7 @@ void Manifold::Impl::FlattenFaces() {
   newHalfedge2.resize(startFace);
   halfedge_ = std::move(newHalfedge2);
 
-  Face2Tri(faceEdge2, halfedgeRef);
+  Face2Tri(faceEdge2, halfedgeRef, faceEdgeStart);
   Finish();
 }
 
diff --git i/src/impl.h w/src/impl.h
index 8f9f537f..5d1c765f 100644
--- i/src/impl.h
+++ w/src/impl.h
@@ -323,7 +323,7 @@ struct Manifold::Impl {
   void GatherFaces(const Impl& old, const Vec<int>& faceNew2Old);
 
   // face_op.cpp
-  void Face2Tri(const Vec<int>& faceEdge, const Vec<TriRef>& halfedgeRef);
+  void Face2Tri(const Vec<int>& faceEdge, const Vec<TriRef>& halfedgeRef, const Vec<int>& faceEdgeStart);
   void FlattenFaces();
   Polygons Slice(double height) const;
   Polygons Project() const;

I don't push it because it is too ugly :P

…cimator

elalish · 2025-03-03T09:07:04Z

Okay, how's this? It's 2x faster than the old decimator, so that's something at least. I still need to see if I can make these hexagons to triangulate nicer.

codecov · 2025-03-03T09:21:51Z

Codecov Report

Attention: Patch coverage is 86.77686% with 16 lines in your changes missing coverage. Please review.

Project coverage is 91.62%. Comparing base (0451671) to head (80c90b4).
Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
src/impl.h	33.33%	8 Missing ⚠️
src/face_op.cpp	93.93%	6 Missing ⚠️
src/impl.cpp	33.33%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1175      +/-   ##
==========================================
+ Coverage   91.60%   91.62%   +0.02%     
==========================================
  Files          32       32              
  Lines        5989     6067      +78     
==========================================
+ Hits         5486     5559      +73     
- Misses        503      508       +5

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

pca006132 · 2025-03-04T06:23:39Z

I think the black formatter version may be automatically updated, causing the format CI error.

pca006132 · 2025-03-04T11:07:23Z

src/edge_op.cpp

@@ -235,6 +235,9 @@ void Manifold::Impl::CleanupTopology() {
 * meshes, thus decreasing the Genus(). It only increases when meshes that have
 * collapsed to just a pair of triangles are removed entirely.
 *
+ * Verts with index less than firstNewVert will be left uncollapsed. This is
+ * zero by default so that everything can be collapsed.


When do we use this zero default? We should probably explain this mesh simplification behavior in the public documentation as well.

Everywhere except the Boolean - mesh ingestion, AsOriginal, etc. Hmm, come to think of it, the issue of decimating input planar geometry that doesn't intersect anything isn't really fixed, since we're still doing it on mesh import and other places. Perhaps those should all be restricted to short edge collapse? I think I'll do this in a separate PR.

Sounds good.

pca006132 · 2025-03-04T11:35:42Z

src/face_op.cpp

-      thisEdge = startEdge;
-      polys.push_back({});
+  if (numFace < 4) {
+    MarkFailure(Error::NoError);


Should we change the name of this function? Done or something? Otherwise it is a bit weird to mark something as failure even though it has no error.

Yeah, I was thinking the same thing. MarkEmpty sounds good.

pca006132 · 2025-03-04T11:38:01Z

src/face_op.cpp

+
+    const int numEdge = polys2.size();
+    if (numEdge < 3) return;
+    const uint64_t inc = (1LL << 32) + numEdge;


We should probably document this so that whoever looks at this later can figure out the pair encoding at a glance. Or maybe just make a helper function because we use this below as well.

pca006132 · 2025-03-04T11:41:49Z

src/polygon.cpp

@@ -936,16 +940,21 @@ namespace manifold {
 * references back to the original vertices.
 * @param epsilon The value of &epsilon;, bounding the uncertainty of the
 * input.
+ * @param allowConvex If true (default), the triangulator will use a fast


Where do we use this?

This was how I fixed Simplify() - the surface looks like garbage with convex triangulation (almost the same as the edge collapse version); ear clipping sorts by quality so it does quite well. Figured others might need it too, so probably best in the public API anyway.

elalish added 4 commits February 21, 2025 21:52

merging

b6cceec

mostly working

0402a32

implemented, but crashes

a578f80

runs serially

8bddd56

elalish requested a review from pca006132 February 28, 2025 20:58

elalish commented Feb 28, 2025

View reviewed changes

pca006132 reviewed Mar 1, 2025

View reviewed changes

src/face_op.cpp Show resolved Hide resolved

src/face_op.cpp Outdated Show resolved Hide resolved

fix compile issues

ca0868c

elalish added 3 commits March 2, 2025 17:28

fixed empty meshes

44111ae

fix parallel error dumping

498b169

Merge branch 'newDecimator' of github.com:elalish/manifold into newDe…

d54feea

…cimator

elalish commented Mar 2, 2025

View reviewed changes

Merge branch 'master' into newDecimator

350fe42

elalish added 2 commits March 3, 2025 22:04

fixed parallelization

2884adb

Merge branch 'newDecimator' of github.com:elalish/manifold into newDe…

716d017

…cimator

fix compiler

c154b4b

elalish added 2 commits March 4, 2025 21:28

run black

959c207

improved triangulation

0e0157c

elalish requested a review from pca006132 March 4, 2025 09:47

pca006132 reviewed Mar 4, 2025

View reviewed changes

elalish added 2 commits March 5, 2025 23:09

remove unreferenced verts

aa0b519

addressing feedback

80c90b4

pca006132 approved these changes Mar 5, 2025

View reviewed changes

elalish merged commit 954df79 into master Mar 6, 2025
27 checks passed

elalish deleted the newDecimator branch March 6, 2025 04:42

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

New decimator #1175

New decimator #1175

elalish commented Feb 28, 2025 •

edited

Loading

elalish Feb 28, 2025

pca006132 left a comment

pca006132 commented Mar 1, 2025

elalish commented Mar 1, 2025

elalish Mar 2, 2025

pca006132 Mar 2, 2025

pca006132 Mar 2, 2025

pca006132 commented Mar 2, 2025

elalish commented Mar 3, 2025 •

edited

Loading

codecov bot commented Mar 3, 2025 •

edited

Loading

pca006132 commented Mar 4, 2025

pca006132 Mar 4, 2025

elalish Mar 5, 2025

pca006132 Mar 5, 2025

pca006132 Mar 4, 2025

elalish Mar 5, 2025

pca006132 Mar 4, 2025

pca006132 Mar 4, 2025

elalish Mar 5, 2025 •

edited

Loading

New decimator #1175

New decimator #1175

Conversation

elalish commented Feb 28, 2025 • edited Loading

Choose a reason for hiding this comment

pca006132 left a comment

Choose a reason for hiding this comment

pca006132 commented Mar 1, 2025

elalish commented Mar 1, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

pca006132 commented Mar 2, 2025

elalish commented Mar 3, 2025 • edited Loading

codecov bot commented Mar 3, 2025 • edited Loading

Codecov Report

pca006132 commented Mar 4, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

elalish Mar 5, 2025 • edited Loading

Choose a reason for hiding this comment

elalish commented Feb 28, 2025 •

edited

Loading

elalish commented Mar 3, 2025 •

edited

Loading

codecov bot commented Mar 3, 2025 •

edited

Loading

elalish Mar 5, 2025 •

edited

Loading