Library does not handle rays parallel to axes about bvh HOT 5 CLOSED

This is an example file with all triangle data. Surface is a regular grid mesh.

In leaf function I always return false for this example so that I can count all visits. Epsilon is set to 0 for reproducibility. If you set this to std::numeric_limits<float>::epsilon() + 1ulp traversal works as expected. For now I'll add a small epsilon value until better solution is found.

bvh-issue60.cpp
data.txt
mesh-preview.png

from bvh.

Thank you for fast response and for taking the time to fix the problem.

from bvh.

Here are a couple of things you could try:

• Adding an epsilon to your ray direction/origin,
• Avoiding rays that are very small,
• Using the robust node intersector (IsRobust = true),

If none of those things help, please send me a small reproducing example, and I'll have a look at it.

from bvh.

TLDR:
Use master with the robust intersector.

Long version:

So I did a little investigation and that's what I found:

• You are not setting up the scene properly. Primitives are connected in the wrong way, creating degenerate triangles.
• I fixed that by changing the order in which coordinates are loaded from the stream, and validated it by exporting the result as OBJ and viewing it in Blender.
• With that, the robust intersector was producing fewer traversal steps, but still too much in my opinion, so I dug into the code and fixed a bug where the inverse of the ray direction was computed via safe_inverse instead of a direct inverse.
• The fast intersector returns wrong results here. That's expected. After all, it's using FMA to compute (pos - ray.org) / ray.dir as pos * ray.inv_dir + inv_org_dir (where inv_org_dir = -ray.org * ray.inv_dir). This means that if the direction is zero, then ray.inv_dir == inf and inv_org_dir == -inf and therefore the FMA returns NaN. To solve that problem, in fast mode, I don't compute the inverse exactly, but instead use FLT_MAX as an inverse if the direction is too close to zero. This explains why when the direction is small, the fast intersector returns incorrect results.

from bvh.

Regarding generating an OBJ file and making sure the geometry is correct, here's the relevant code (modified version of your code):

std::fstream stream("data.txt", std::fstream::in);

while (stream) {
    Vec3 a, b, c;
    
    // WRONG ORDER:
    //stream >> a[0] >> c[1] >> c[2]
    //       >> b[0] >> b[1] >> b[2]
    //       >> c[0] >> a[1] >> a[2];
    
    // CORRECT ORDER:
    stream >> a[0] >> a[1] >> a[2]
           >> b[0] >> b[1] >> b[2]
           >> c[0] >> c[1] >> c[2];

    if (stream)
        precomputed_tris.emplace_back(a, b, c);
}

std::ofstream s("data.obj");
int i = 0;
for (auto& precomp_tri : precomputed_tris) {
    auto tri = precomp_tri.convert_to_tri();
    s << "v " << tri.p0[0] << " " << tri.p0[1] << " " << tri.p0[2] << '\n';
    s << "v " << tri.p1[0] << " " << tri.p1[1] << " " << tri.p1[2] << '\n';
    s << "v " << tri.p2[0] << " " << tri.p2[1] << " " << tri.p2[2] << '\n';
    s << "f -1 -2 -3\n";
}

from bvh.