A fast BVH using SAH in rust
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
media
src Fix some lints Aug 23, 2018
.gitignore Initial cargo project Sep 22, 2016
.rustfmt.toml
.travis.yml Fix .travis.yml and allow failures for stable and beta Dec 18, 2016
Cargo.toml This is 0.2.3 Aug 23, 2018
LICENSE Initial commit Sep 22, 2016
README.md Add dependency status badge Aug 23, 2018

README.md

bvh

Build Status Docs Status dependency status license Crates.io Crates.io

A crate which exports rays, axis-aligned bounding boxes, and binary bounding volume hierarchies.

About

This crate can be used for applications which contain intersection computations of rays with primitives. For this purpose a binary tree BVH (Bounding Volume Hierarchy) is of great use if the scene which the ray traverses contains a huge number of primitives. With a BVH the intersection test complexity is reduced from O(n) to O(log2(n)) at the cost of building the BVH once in advance. This technique is especially useful in ray/path tracers. For use in a shader this module also exports a flattening procedure, which allows for iterative traversal of the BVH. This library is built on top of nalgebra.

Example

use bvh::aabb::{AABB, Bounded};
use bvh::bvh::BVH;
use bvh::nalgebra::{Point3, Vector3};
use bvh::ray::Ray;

let origin = Point3::new(0.0,0.0,0.0);
let direction = Vector3::new(1.0,0.0,0.0);
let ray = Ray::new(origin, direction);

struct Sphere {
    position: Point3<f32>,
    radius: f32,
}

impl Bounded for Sphere {
    fn aabb(&self) -> AABB {
        let half_size = Vector3::new(self.radius, self.radius, self.radius);
        let min = self.position - half_size;
        let max = self.position + half_size;
        AABB::with_bounds(min, max)
    }
}

let mut spheres = Vec::new();
for i in 0..1000u32 {
    let position = Point3::new(i as f32, i as f32, i as f32);
    let radius = (i % 10) as f32 + 1.0;
    spheres.push(Sphere {
        position: position,
        radius: radius,
    });
}

let bvh = BVH::build(&spheres);
let hit_sphere_aabbs = bvh.traverse_recursive(&ray, &spheres);

Optimization

This crate provides BVH updating, which is also called optimization. With BVH optimization you can mutate the shapes on which the BVH is built and update the tree accordingly without rebuilding it completely. This method is very useful when there are only very few changes to a huge scene. When the major part of the scene is static, it is faster to update the tree, instead of rebuilding it from scratch.

Drawbacks

First of all, optimizing is not helpful if more than half of the scene is not static. This is due to how optimizing takes place: Given a set of indices of all shapes which have changed, the optimize procedure tries to rotate fixed constellations in search for a better surface area heuristic (SAH) value. This is done recursively from bottom to top while fixing the AABBs in the inner nodes of the BVH. Which is why it is inefficient to update the BVH in comparison to rebuilding, when a lot of shapes have moved.

Another problem with updated BVHs is, that the resulting BVH is not optimal. Assume that the scene is composed of two major groups separated by a large gap. When one shape moves from one group to another, the updating procedure will not be able to find a sequence of bottom-up rotations which inserts the shape deeply into the other branch.

The following benchmarks are run with two different datasets:

  • A randomly generated scene with unit sized cubes containing a total of (1200, 12000, and 120000 triangles).
  • Sponza, a popular scene for benchmarking graphics engines.

Intersection via traversal of the list of triangles

test testbase::bench_intersect_120k_triangles_list                       ... bench:   1,018,566 ns/iter (+/- 91,405)
test testbase::bench_intersect_sponza_list                               ... bench:     669,474 ns/iter (+/- 18,928)

This is the most naive approach to intersecting a scene with a ray. It defines the baseline.

Intersection via traversal of the list of triangles with AABB checks

test testbase::bench_intersect_120k_triangles_list_aabb                  ... bench:     400,877 ns/iter (+/- 23,775)
test testbase::bench_intersect_sponza_list_aabb                          ... bench:     206,014 ns/iter (+/- 4,508)

AABB checks are cheap, compared to triangle-intersection algorithms. Therefore, preceeding AABB checks increase intersection speed by filtering negative results a lot faster.

Build of a BVH from scratch

test bvh::bvh::tests::bench_build_1200_triangles_bvh                     ... bench:   1,300,824 ns/iter (+/- 32,262)
test bvh::bvh::tests::bench_build_12k_triangles_bvh                      ... bench:  15,327,304 ns/iter (+/- 360,985)
test bvh::bvh::tests::bench_build_120k_triangles_bvh                     ... bench: 181,138,173 ns/iter (+/- 5,296,719)

test bvh::bvh::tests::bench_build_sponza_bvh                             ... bench: 120,335,877 ns/iter (+/- 3,787,414)

As you can see, building a BVH takes a long time. Building a BVH is only useful if the number of intersections performed on the scene exceeds the build duration. This is the case in applications such as ray and path tracing, where the minimum number of intersections is 1280 * 720 for an HD image.

Intersection via BVH traversal

test bvh::bvh::tests::bench_intersect_1200_triangles_bvh                 ... bench:         202 ns/iter (+/- 3)
test bvh::bvh::tests::bench_intersect_120k_triangles_bvh                 ... bench:         959 ns/iter (+/- 26)
test bvh::bvh::tests::bench_intersect_12k_triangles_bvh                  ... bench:         461 ns/iter (+/- 14)
test bvh::bvh::tests::bench_intersect_sponza_bvh                         ... bench:       1,784 ns/iter (+/- 202)

These performance measurements show that traversing a BVH is much faster than traversing a list.

Optimization

The benchmarks for how long it takes to update the scene also contain a randomization process which takes some time.

test bvh::optimization::tests::bench_randomize_120k_50p                  ... bench:  14,248,069 ns/iter (+/- 2,368,251)

test bvh::optimization::tests::bench_optimize_bvh_120k_00p               ... bench:   2,338,563 ns/iter (+/- 59,248)
test bvh::optimization::tests::bench_optimize_bvh_120k_01p               ... bench:  12,690,322 ns/iter (+/- 5,235,405)
test bvh::optimization::tests::bench_optimize_bvh_120k_10p               ... bench: 117,318,325 ns/iter (+/- 34,879,930)
test bvh::optimization::tests::bench_optimize_bvh_120k_50p               ... bench: 502,788,600 ns/iter (+/- 161,281,887)

// TODO Sponza benchmarks.

This is the place where you have to differentiate between rebuilding the tree from scratch or trying to optimize the old one. These tests show the impact of moving around a particular percentage of shapes (10p => 10%). It is important to note that the randomization process here moves triangles around indiscriminately. This will also lead to cases where the BVH would have to be restructured completely.

Intersection after the optimization

These intersection tests are grouped by dataset and by the BVH generation method.

  • _after_optimize uses a BVH which was kept up to date with calls to optimize, while
  • _with_rebuild uses the same triangle data as _after_optimize, but constructs a BVH from scratch.

120K Triangles

test bvh::optimization::tests::bench_intersect_120k_after_optimize_00p   ... bench:         968 ns/iter (+/- 31)
test bvh::optimization::tests::bench_intersect_120k_after_optimize_01p   ... bench:     147,160 ns/iter (+/- 12,886)
test bvh::optimization::tests::bench_intersect_120k_after_optimize_10p   ... bench:   1,624,675 ns/iter (+/- 758,933)
test bvh::optimization::tests::bench_intersect_120k_after_optimize_50p   ... bench:   2,775,067 ns/iter (+/- 751,818)

test bvh::optimization::tests::bench_intersect_120k_with_rebuild_00p     ... bench:         964 ns/iter (+/- 38)
test bvh::optimization::tests::bench_intersect_120k_with_rebuild_01p     ... bench:       1,016 ns/iter (+/- 16)
test bvh::optimization::tests::bench_intersect_120k_with_rebuild_10p     ... bench:       2,025 ns/iter (+/- 213)
test bvh::optimization::tests::bench_intersect_120k_with_rebuild_50p     ... bench:       2,373 ns/iter (+/- 251)

Sponza

test bvh::optimization::tests::bench_intersect_sponza_after_optimize_00p ... bench:       1,824 ns/iter (+/- 114)
test bvh::optimization::tests::bench_intersect_sponza_after_optimize_01p ... bench:       3,791 ns/iter (+/- 308)
test bvh::optimization::tests::bench_intersect_sponza_after_optimize_10p ... bench:       4,794 ns/iter (+/- 212)
test bvh::optimization::tests::bench_intersect_sponza_after_optimize_50p ... bench:       7,492 ns/iter (+/- 807)

test bvh::optimization::tests::bench_intersect_sponza_with_rebuild_00p   ... bench:       1,823 ns/iter (+/- 145)
test bvh::optimization::tests::bench_intersect_sponza_with_rebuild_01p   ... bench:       1,957 ns/iter (+/- 114)
test bvh::optimization::tests::bench_intersect_sponza_with_rebuild_10p   ... bench:       2,414 ns/iter (+/- 209)
test bvh::optimization::tests::bench_intersect_sponza_with_rebuild_50p   ... bench:       3,135 ns/iter (+/- 322)

This set of tests shows the impact of randomly moving triangles around and producing degenerated trees. The 120K Triangles dataset has been updated randomly. The Sponza scene was updated using a method which has a maximum offset distance for shapes. This simulates a more realistic scenario.

We also see that the Sponza scene by itself contains some structures which can be tightly wrapped in a BVH. By mowing those structures around we destroy the locality of the triangle groups which leads to more branches in the BVH requiring a check, thus leading to a higher intersection duration.