Feeding a better initial penetration axis to GJK this improves performance

[jrouwe]

No description provided.

[sonarcloud]

Kudos, SonarCloud Quality Gate passed!   

0 Bugs
0 Vulnerabilities
0 Security Hotspots
0 Code Smells

100.0% Coverage
0.0% Duplication

[virt00l]

Sorry for necroposting, but this change seems to cause this issue:

// Don't supply a zero ioV, we only want to get points on the hull of the Minkowsky sum and not internal
JPH_ASSERT(!ioV.IsNearZero());

actual data from game mesh (triangle seems to be long, but not degenerate?):

v0 = array((0.279280782, -0.467894495, 0.105353981))
v1 = array((0.280091137, -0.467545986, 0.104968153))
v2 = array((0.277769983, -0.469322920, 0.105590254))
n = np.cross(v1 - v0, v2 - v0)
print(v1 - v0)
print(v2 - v0)
print(n)
print(n.dot(n))
...
[ 0.00081035  0.00034851 -0.00038583]
[-0.0015108  -0.00142842  0.00023627]
[ -4.68783094e-07   3.91443550e-07  -6.31004292e-07]
7.71152058489e-13

Not sure how to better handle this.
Check length and fallback to sAxisX instead of using normal ? Wouldn't this length check + branch cause net loss of said 5%?
Revert this change? Increase threshold (e.g. 1e-12 -> 1e-14)? Fix the data (mesh)?

[jrouwe]

Hello,

Is it just this assert that is triggering or is there an actual issue? What is the full call stack? Is this a heavily scaled MeshShape perhaps?

In any case, the triangle is really really tiny (less than 1 mm^2), and it looks like it is barely bigger than the degenerate check that would have thrown the triangle away:

JoltPhysics/Jolt/Physics/Collision/Shape/MeshShape.cpp

Line 98 in 153291b

if (tri.IsDegenerate(mTriangleVertices) // Degenerate triangle

Are you able to clean up your mesh before providing it to Jolt? (it's just not a good idea to put triangles this small into the collision detection pipeline)

[virt00l]

Hello,

Is it just this assert that is triggering or is there an actual issue?

Well, I havent considered that there could be the difference. Will indeed try to comment it and see what happen

What is the full call stack

>	enlisted-dev.exe!JPH::CollideConvexVsTriangles::Collide(JPH::Vec3 inV0, JPH::Vec3 inV1, JPH::Vec3 inV2, unsigned char inActiveEdges, const JPH::SubShapeID & inSubShapeID2) Line 82	C++	Symbols loaded.
 	[Inline Frame] enlisted-dev.exe!JPH::MeshShape::sCollideConvexVsMesh::Visitor::VisitTriangle(JPH::Vec3 inV0, JPH::Vec3 inV1, JPH::Vec3 inV2, unsigned char inActiveEdges, JPH::SubShapeID inSubShapeID2) Line 1091	C++	Symbols loaded.
 	[Inline Frame] enlisted-dev.exe!JPH::MeshShape::WalkTreePerTriangle<Visitor>::ChainedVisitor::VisitTriangles(const JPH::TriangleCodecIndexed8BitPackSOA4Flags::DecodingContext & ioContext, const void * inTriangles, int inNumTriangles, unsigned int inTriangleBlockID) Line 487	C++	Symbols loaded.
 	[Inline Frame] enlisted-dev.exe!JPH::NodeCodecQuadTreeHalfFloat<1>::DecodingContext::WalkTree(const unsigned char * inBufferStart, const JPH::TriangleCodecIndexed8BitPackSOA4Flags::DecodingContext & inTriangleContext, JPH::MeshShape::WalkTreePerTriangle<Visitor>::ChainedVisitor & ioVisitor) Line 260	C++	Symbols loaded.
 	[Inline Frame] enlisted-dev.exe!JPH::MeshShape::WalkTree(JPH::MeshShape::WalkTreePerTriangle<Visitor>::ChainedVisitor & ioVisitor) Line 437	C++	Symbols loaded.
 	[Inline Frame] enlisted-dev.exe!JPH::MeshShape::WalkTreePerTriangle(const JPH::SubShapeIDCreator & inSubShapeIDCreator2, JPH::MeshShape::sCollideConvexVsMesh::Visitor & ioVisitor) Line 501	C++	Symbols loaded.
 	enlisted-dev.exe!JPH::MeshShape::sCollideConvexVsMesh(const JPH::Shape * inShape1, const JPH::Shape * inShape2, JPH::Vec3 inScale1, JPH::Vec3 inScale2, const JPH::Mat44 & inCenterOfMassTransform1, const JPH::Mat44 & inCenterOfMassTransform2, const JPH::SubShapeIDCreator & inSubShapeIDCreator1, const JPH::SubShapeIDCreator & inSubShapeIDCreator2, const JPH::CollideShapeSettings & inCollideShapeSettings, JPH::CollisionCollector<JPH::CollideShapeResult,JPH::CollisionCollectorTraitsCollideShape> & ioCollector, const JPH::ShapeFilter & inShapeFilter) Line 1096	C++	Symbols loaded.

In any case, the triangle is really really tiny (less than 1 mm^2), and it looks like it is barely bigger than the degenerate check >that would have thrown the triangle away:

I forgot to mention - this coords are in after this translation mTransform2To1 * (mScale2 * inV0), it's scaled, but not that much (0.6875), values are taken raw from debugger (last stack frame of above stack trace):

triangle:
{mValue=0x000000a8e5a39c60 {0.279185504, -0.478402555, -0.0955970585, 1.00000000}
{mValue=0x000000a8e5a39c70 {0.278547287, -0.478054047, -0.0962281227, 1.00000000}
{mValue=0x000000a8e5a39c80 {0.279921055, -0.479830980, -0.0942564011, 1.00000000}

mTransform2To1:
{-0.854890704, 0.00000000, -0.518808305, 0.00000000}
{0.00000000, 1.00000000, 0.00000000, 0.00000000}
{0.518808305, 0.00000000, -0.854890704, 0.00000000}
{0.286158055, -1.08555472, -0.273416281, 1.00000000}

From first look into the coll mesh it doesn't look too bad, but I'm currently in process of gathering additional info

[virt00l]

Okay, we found the root cause of it (it seems) - it was broken mesh after all (was fixed recently, thats why I wasnt be able spot it right away in asset viewer, but existed in slightly outdated form within game)