shinjiogaki / bvh Goto Github PK

以下の資料の P53 にある「C++は違う型のunionが許されない」という記載について、元ソースを見た限り「union で代入したメンバとは違うメンバから値を読み取った場合、値の保証ができない」と言っているだけなので、動作自体はする（資料の書き方でもコンパイルは通る）と思いますが、いかがでしょう？

https://github.com/shinjiogaki/bvh/blob/4250ba1/CEDEC2019.pdf

Hello,
I have randomly underexpanded bboxes if i use
bvh_binary.cpp: 73 std::for_each(std::execution::par, leaves.begin(), leaves.end(), [&](glm::uvec3 &leaf)
these two screenshots from two different launches of the same build

however if i change the execution policy to sequential (std::execution::seq) everything is ok, the problem mainly manifests itself on low-poly inputs. Resulting tree is valid in both cases as i can see, the problem only in bbox calculation. I ran this on 112 threads workstation.

Also my code for test intersection if any, counts hits and outputs as brightness. Could the problem be on my side ?

inline int TEST_intersect_bvh_ray(int  ndIdx, const  TEST :: LBVH& acc, const triangle* tris, const vec3 ro, const vec3 rd)
	{
		const auto * nd = &acc.Nodes[ndIdx];
		int nbHits = 0;
		float tnear, tfar;
		if (intersect(aabb(&nd->Box.Min.x, &nd->Box.Max.x), ro, rd, tnear, tfar) == false)
		 	return 0;
 
		if ((nd->L & (1))) // is leaf ?
		{
			vec3 uvw;
			float  tmax = _FLT_MAX;
			if (intersect_triangle_ray(tris[(nd->L - 1) / 2], ro, rd, uvw, tmax))
				nbHits++;
		}
		else
		{
			nbHits += TEST_intersect_bvh_ray(nd->L / 2, acc, tris, ro, rd);
		}

		if ((nd->R & (1))) // is leaf ?
		{
			vec3 uvw;
			float  tmax = _FLT_MAX;
			if (intersect_triangle_ray(tris[(nd->R - 1) / 2], ro, rd, uvw, tmax))
				nbHits++;
		}
		else
		{
			nbHits += TEST_intersect_bvh_ray(nd->R / 2, acc, tris, ro, rd);
		}

		return nbHits;
	}

bvh_binary.cpp
line 104

// leaf/node index
const auto index = is_leaf ? leaves[current].x * 2 + 1 : current * 2;

I don't understand this line.
Is the "*2" a left over from the 2d version?

Why not just something like that?

// leaf/node index
const auto index = is_leaf ? current +1 : current ;