[TOC]
今天遇到一个问题,角色卡在一个模型边上,在PVD看模型也比较正常,一直调试跟踪,最终找到了问题所在。
原因
1.问题出在最后一个sweepCapsuleTriangles函数里面,代码如下:
#ifdef __SPU__
Vec3V dummy1, dummy2;
FloatV result = distanceSegmentTriangleSquared(
V3LoadU(capsule.p0), V3LoadU(capsule.p1), V3LoadU(p0), V3LoadU(p1), V3LoadU(p2), dummy1, dummy2);
PxReal dist2 = FStore(result);
#else
// AP: switching to SIMD version produced -25% regression in Sweep*SphereHfld
PxReal dist2 = distanceSegmentTriangleSquared(capsule.p0, segmentExtent, p0, p1 - p0, p2 - p0);
#endif
if (dist2<=r2)
{
hitIndex = i;
t = 0.0f;
normal = -unitDir;
outFlags = PxHitFlag::eDISTANCE|PxHitFlag::eNORMAL;
_triNormal = denormalizedNormal.getNormalized();
return true;
}
当p0、p1和p2有有两个点相同的时候,会导致p1-p0和p2-p0其中一个向量为0,得到的dist2为负无穷大,所以就会执行代码中的条件语句。最终结果是move函数移动成功了,但是只移动了0的位移,表现上就是一直卡在那个地方。
- 至于顶点一样的情况,其实并不是完全一样,而是有很小的差异。再加上浮点数运算会丢失精度,这个很小的差异被抹掉了,所以到上面代码的时候就出现相同的点了。
计算的代码位置(没看太明白):
PX_FORCE_INLINE void Gu::TriangleMesh::computeWorldTriangle(PxTriangle& worldTri, PxTriangleID triangleIndex, const Cm::Matrix34& worldMatrix, PxU32* PX_RESTRICT vertexIndices, PxU32* PX_RESTRICT adjacencyIndices) const
{
PxU32 vref0, vref1, vref2;
if(mMesh.has16BitIndices())
{
const Gu::TriangleT<PxU16>& T = ((const Gu::TriangleT<PxU16>*)getTrianglesFast())[triangleIndex];
vref0 = T.v[0];
vref1 = T.v[1];
vref2 = T.v[2];
}
else
{
const Gu::TriangleT<PxU32>& T = ((const Gu::TriangleT<PxU32>*)getTrianglesFast())[triangleIndex];
vref0 = T.v[0];
vref1 = T.v[1];
vref2 = T.v[2];
}
const PxVec3* PX_RESTRICT vertices = getVerticesFast();
worldTri.verts[0] = worldMatrix.transform(vertices[vref0]);
worldTri.verts[1] = worldMatrix.transform(vertices[vref1]);
worldTri.verts[2] = worldMatrix.transform(vertices[vref2]);
if(vertexIndices)
{
vertexIndices[0] = vref0;
vertexIndices[1] = vref1;
vertexIndices[2] = vref2;
}
if(adjacencyIndices)
{
if(mMesh.getAdjacencies())
{
adjacencyIndices[0] = mMesh.getAdjacencies()[triangleIndex*3 + 0];
adjacencyIndices[1] = mMesh.getAdjacencies()[triangleIndex*3 + 1];
adjacencyIndices[2] = mMesh.getAdjacencies()[triangleIndex*3 + 2];
}
else
{
adjacencyIndices[0] = 0xffffffff;
adjacencyIndices[1] = 0xffffffff;
adjacencyIndices[2] = 0xffffffff;
}
}
}
- 问题数据
出现问题的时候,断点的数据如下:
vertices[vref0] = {x = -4.10000086, y = -0.200000167, z = -3.56512594}
vertices[vref1] = {x = -4.10000086, y = -0.200000077, z = -3.56512594}
vertices[vref2] = {x = -4.10000086, y = -0.200000212, z = -4.10000229}
worldMatrix.transform.base3 = {x=72.3379974 y=4.76143503 z=18.5843773 }
// 得到这个三角形前两个点完全一样
worldTri.verts[0] = { x = 68.237990, y = 4.56143475, z = 15.0192509}
worldTri.verts[1] = { x = 68.237990, y = 4.56143475, z = 15.0192509}
worldTri.verts[2] = { x = 68.237990, y = 4.56143475, z = 14.4843750}
简单来说就是:
4.76143503 + -0.200000167 = 4.56143475
4.76143503 + -0.200000077 = 4.56143475
至于为什么请参考另外一篇文章[1]为什单精度浮点数的有效位时7位
解决
- 1.把顶点传到PhysX之前,先检查一遍。可以删掉重新组织Vertex和Index,我简单处理的方法是把两个顶点很近的时候,标记为同一个索引,这样的三角形比较容易标识
if (fMinDist <= 1e-5) { // set the same index, and then cooking will process it. eREMOVE_DUPLICATED_TRIANGLES pIndices[3 * i + nIndexIndex1] = pIndices[3 * i + nIndexIndex2]; }
- 2.PhysxCooking的参数增加去重的Flag,这样上面的三角形就会被PhysX剔除,从而就解决了这样的问题:如下
physx::PxCookingParams cookingParams(scale);
cookingParams.buildTriangleAdjacencies = true;
cookingParams.meshPreprocessParams = PxMeshPreprocessingFlags(PxMeshPreprocessingFlag::eWELD_VERTICES |
PxMeshPreprocessingFlag::eREMOVE_UNREFERENCED_VERTICES | PxMeshPreprocessingFlag::eREMOVE_DUPLICATED_TRIANGLES);