Instances about bvh

[madmann91]

You can update the bounding boxes of every node in a for loop like so:

// It's important to process the nodes in reverse order, as this makes
// sure that children are processed before their parents.
for (auto& node : std::ranges::reverse_view { bvh.nodes }) {
  if (node.is_leaf()) {
    // refit node according to contents
    ...
  } else {
    auto& left  = bvh.nodes[node.index.first_id];
    auto& right = bvh.nodes[node.index.first_id + 1];
    node.set_bbox(left.get_bbox().extend(right.get_bbox()));
  }
}

I should probably add a method for this. Anyway, this method is very fast but should not be run in parallel as it expects the children to have been processed before the parents. The method used in the previous version of the repository works in parallel but is only faster when the BVH is extremely large.

[madmann91]

Yes, they are still supported. The interface is a bit lower-level in the new version, so you only need an axis-aligned bounding box and a center point for each primitive you need to insert in the BVH. For you, this means something like:

std::vector<BBox> bboxes(obbs.size());
std::vector<Vec3> centers(obbs.size());
executor.for_each(0, tris.size(), [&] (size_t begin, size_t end) {
    for (size_t i = begin; i < end; ++i) {
        bboxes[i]  = get_box(obbs[i]);
        centers[i] = obbs[i].pos;
    }
});

typename bvh::v2::DefaultBuilder<Node>::Config config;
config.quality = bvh::v2::DefaultBuilder<Node>::Quality::High;
auto bvh = bvh::v2::DefaultBuilder<Node>::build(thread_pool, bboxes, centers, config);

You can get the AABB of an OBB this way (you do not need to transform the 8 points of a 3D OBB):

// Assuming the corners are obtained by multiplying the matrix with (+-1, +-1, +-1)
struct OBB {
    Mat3x3 mat;
    Vec3 pos;
};

inline Box3 get_bbox(const OBB& obb) {
    auto abs_mat = abs(obb.mat);
    auto half_extents = abs_mat * Vec3 { 1, 1, 1 };
    return Box3 { obb.pos - half_extents, obb.pos + half_extents };
};

Thanks, that works! But what about updates/refitting? In version 1 I could use the HierarchyRefitter, but I see no equivalent in version 2. I create the bvh like you describe:

bvh = ::bvh::v2::DefaultBuilder<pragma::bvh::Node>::build(GetThreadPool(), bboxes, centers, config);

After creation I need to update the bboxes/centers every frame, but I see no way to access those after the bvh has been built. Can the ReinsertionOptimizer be used for refitting somehow?

::bvh::v2::ReinsertionOptimizer<Node>::optimize(GetThreadPool(), bvh);

There is now a refitting method which allows you to refit the BVH serially (since #74 has been merged). I might add a parallel version later.

[madmann91]

Hello! You may take a look at the example on custom primitives. An instance is just a primitive like any other, which itself contains a BVH and a matrix. Before intersecting the instance, you transform the ray by the inverse of the matrix. That's basically it.

[Silverlan]

Hello! You may take a look at the example on custom primitives. An instance is just a primitive like any other, which itself contains a BVH and a matrix. Before intersecting the instance, you transform the ray by the inverse of the matrix. That's basically it.

Are custom primitives still supported? I noticed that the example for them is no longer available, and I need a BVH with OBBs as primitives.

[madmann91]

Yes, they are still supported. The interface is a bit lower-level in the new version, so you only need an axis-aligned bounding box and a center point for each primitive you need to insert in the BVH. For you, this means something like:

std::vector<BBox> bboxes(obbs.size());
std::vector<Vec3> centers(obbs.size());
executor.for_each(0, tris.size(), [&] (size_t begin, size_t end) {
    for (size_t i = begin; i < end; ++i) {
        bboxes[i]  = get_box(obbs[i]);
        centers[i] = obbs[i].pos;
    }
});

typename bvh::v2::DefaultBuilder<Node>::Config config;
config.quality = bvh::v2::DefaultBuilder<Node>::Quality::High;
auto bvh = bvh::v2::DefaultBuilder<Node>::build(thread_pool, bboxes, centers, config);

You can get the AABB of an OBB this way (you do not need to transform the 8 points of a 3D OBB):

// Assuming the corners are obtained by multiplying the matrix with (+-1, +-1, +-1)
struct OBB {
    Mat3x3 mat;
    Vec3 pos;
};

inline Box3 get_bbox(const OBB& obb) {
    auto abs_mat = abs(obb.mat);
    auto half_extents = abs_mat * Vec3 { 1, 1, 1 };
    return Box3 { obb.pos - half_extents, obb.pos + half_extents };
};

[Silverlan]

Yes, they are still supported. The interface is a bit lower-level in the new version, so you only need an axis-aligned bounding box and a center point for each primitive you need to insert in the BVH. For you, this means something like:

std::vector<BBox> bboxes(obbs.size());
std::vector<Vec3> centers(obbs.size());
executor.for_each(0, tris.size(), [&] (size_t begin, size_t end) {
    for (size_t i = begin; i < end; ++i) {
        bboxes[i]  = get_box(obbs[i]);
        centers[i] = obbs[i].pos;
    }
});

typename bvh::v2::DefaultBuilder<Node>::Config config;
config.quality = bvh::v2::DefaultBuilder<Node>::Quality::High;
auto bvh = bvh::v2::DefaultBuilder<Node>::build(thread_pool, bboxes, centers, config);

You can get the AABB of an OBB this way (you do not need to transform the 8 points of a 3D OBB):

// Assuming the corners are obtained by multiplying the matrix with (+-1, +-1, +-1)
struct OBB {
    Mat3x3 mat;
    Vec3 pos;
};

inline Box3 get_bbox(const OBB& obb) {
    auto abs_mat = abs(obb.mat);
    auto half_extents = abs_mat * Vec3 { 1, 1, 1 };
    return Box3 { obb.pos - half_extents, obb.pos + half_extents };
};

Thanks, that works! But what about updates/refitting? In version 1 I could use the HierarchyRefitter, but I see no equivalent in version 2. I create the bvh like you describe:

bvh = ::bvh::v2::DefaultBuilder<pragma::bvh::Node>::build(GetThreadPool(), bboxes, centers, config);

After creation I need to update the bboxes/centers every frame, but I see no way to access those after the bvh has been built. Can the ReinsertionOptimizer be used for refitting somehow?

::bvh::v2::ReinsertionOptimizer<Node>::optimize(GetThreadPool(), bvh);

[tigrazone]

I need example for use in vulkan, for example.
Bvh can be create on CPU but usage on GPU

[madmann91]

I need example for use in vulkan, for example. Bvh can be create on CPU but usage on GPU

Just look at the traversal code in bvh.h. You can also look at https://madmann91.github.io/2021/01/06/bvhs-part-2.html for an introduction to BVH traversal. This article contains implementation code that you can just port to your target language (GLSL/HLSL).