mesh->csg stack overflow

[Jack Rusher]

(-> (cuboid 80)
    (g/as-mesh {:mesh (gm/gmesh)})
    csg/mesh->csg)

... works as expected, but if I try to catmull-clark the mesh before converting to CSG, like this:

(-> (cuboid 80)
    (g/as-mesh {:mesh (gm/gmesh)})
    sd/catmull-clark
    csg/mesh->csg)

... I get this stack overflow:

                       RT.java:  721  clojure.lang.RT/get
                      csg.cljc:   58  thi.ng.geom.mesh.csg$split_poly/invoke
                      csg.cljc:  143  thi.ng.geom.mesh.csg$csg_node$fn__18902/invoke
         PersistentVector.java:  333  clojure.lang.PersistentVector/reduce
                      core.clj: 6518  clojure.core/reduce
                      csg.cljc:  142  thi.ng.geom.mesh.csg$csg_node/invoke

... do more complicated meshes push the recursive csg-node function too deep?

[Karsten Schmidt]

Hi @jackrusher - there definitely is a point where a complex mesh will blow the stack, but it shouldn't be after 1 iteration of subdivision. Will investigate! Thanks

[Karsten Schmidt]

So, the thing works if you pre-tessellate the meshes. This makes me think that there're might be some issues with face normals in the original quad mesh. Because the face normals are used to define cutting planes in the BSP construction, if there're invalid values (e.g. two neighboring faces with opposite orientations), then the BSP construction will go into an infinite loop and blow the stack...

(require '[thi.ng.geom.cuboid :as c])
(require '[thi.ng.geom.gmesh :as gm])
(require '[thi.ng.geom.core :as g])
(require '[thi.ng.geom.mesh.csg :as csg])
(require '[thi.ng.geom.mesh.io :as mio])
(require '[clojure.java.io :as io])

(def m (-> (c/cuboid 80) (g/as-mesh {:mesh (gm/gmesh)})))
(def m2 (sd/catmull-clark m))
(def m3 (last (take 4 (iterate sd/catmull-clark (g/translate m [60 0 0])))))
(def m4 (apply csg/union (map (comp csg/mesh->csg g/tessellate) [m2 m3])))
(with-open [o (io/output-stream "foo.ply")]
  (mio/write-ply (mio/wrapped-output-stream o) (csg/csg->mesh m4)))

[Karsten Schmidt]

Btw. The 4 iterations of subdivs seem to be the maximum for this example. With 5 it throws a different SO, but this time it has to do with concat's laziness...

#error {
 :cause nil
 :via
 [{:type clojure.lang.Compiler$CompilerException
   :message "java.lang.StackOverflowError, compiling:(form-init8102643818965635375.clj:1:9)"
   :at [clojure.lang.Compiler$InvokeExpr eval "Compiler.java" 3628]}
  {:type java.lang.StackOverflowError
   :message nil
   :at [clojure.core$seq__4128 invoke "core.clj" 137]}]
 :trace
 [[clojure.core$seq__4128 invoke "core.clj" 137]
  [clojure.core$concat$fn__4215 invoke "core.clj" 691]
  [clojure.lang.LazySeq sval "LazySeq.java" 40]
  [clojure.lang.LazySeq seq "LazySeq.java" 49]
  [clojure.lang.RT seq "RT.java" 507]
  [clojure.core$seq__4128 invoke "core.clj" 137]
  [clojure.core$concat$cat__4217$fn__4218 invoke "core.clj" 700]
  [clojure.lang.LazySeq sval "LazySeq.java" 40]
  [clojure.lang.LazySeq seq "LazySeq.java" 56]
  [clojure.lang.RT seq "RT.java" 507]
 ...

[Karsten Schmidt]

Proposed fix: Avoid concat calls in csg/clip-polygons & csg/all-polygons, use into or reducers' cat

[Jack Rusher]

Thanks, mate. Having fun with CSG now!

[Karsten Schmidt]

Great! :) Looking forward to see some of that - alas due to the current major mesh refactoring, there's no point in attempting to fix the above in master before the other things have settled