Hierarchical Bitmask Implicit Grids for Efficient Point-in-Volume Queries on the GPU

Implicit grids are a practical data structure for world-space, scene-wide spatial lookups. They can be built in sub-millisecond times for hundreds of thousands of objects and queried very efficiently on the GPU, making them suitable for light culling and decal rendering, especially when typical solutions like deferred & clustered shading are not available.

The pre-print of our paper is available here. This program was used for the benchmarks in the paper.

Code overview

Despite being by the same group that makes Tauray, it is not based on it due to differences in goals. This is a gutted version of an unrelated codebase that may be open-sourced in full at some point later on.

Files related to implicit grid implementation:

Name	Purpose
gfx/clustering.glsl	Lookups from the data structure
gfx/clustering.comp	Builds implicit grid
gfx/clustering_hierarchy.comp	Builds hierarchy layer
gfx/clustering_stage.{cc,hh}	CPU-side render pass code
gfx/decal_ranges.comp	Evaluates range of decal volume across axes
gfx/light_ranges.comp	Evaluates range of light volume across axes
gfx/decal_order.comp	Calculates sorting key for decals
gfx/light_morton.comp	Calculates sorting key for lights

The clustering term is used only for historical reasons; binning would be better.

Building

The program has been tested only on Ubuntu 22.04, don't expect a Windows build to be easy.

cmake -S . -B release -DCMAKE_BUILD_TYPE=Release
cmake --build release

Getting the expected scene

We don't include the Sponza scene in the repo, you'll need to download that separately. You can get the .gltf version from here, import it into Blender and export it as a .glb file, which is expected by this program. Place it into data/sponza.glb.

Running

release/ig_benchmark

The program will start up as a black window.

Initially, it's set up to benchmark ray traced light culling. By pressing 'f', you can swap between light culling and decal rendering tests. By pressing 'r', you can swap between ray tracing and multi-view tests.

You can start the currently selected benchmark with the 't' key. Timing data is printed in stdout.