arbeiter / compute_rasterizer

Rendering Point Clouds with Compute Shaders

About

This repository contains the source code for our papers about real-time software rasterization of point clouds, which can be 10 to 100 times faster than GL_POINTS. This is possible because GL_POINTS is built upon the triangle-oriented rendering pipeline that is not optimal for pixel-sized points.

The basic idea is to spawn a compute shader that transforms points to screen space, encodes depth and color into a single 64 bit integer, and uses atomicMin to compute the closest point for each pixel. The color value is then extracted from the interleaved depth+color buffer and converted into a regular OpenGL texture for display.

The latest improvement also groups about 10k points into batches, and each compute workgroup(128 threads) renders a batch(10k points), i.e., each thread renders about 80 points. This allows several batch-level optimizations such as frustum culling, LOD rendering, and adaptive precision. Adaptive precision picks a sufficient coordinate precision (typically just 10 bit per axis) depending on the projected batch size, which boosts brute-force performance due to lower memory bandwidth requirements.

The main branch is a slightly more user friendly version that allows loading LAS files via drag&drop. Other branches contain snapshots of the code made after evaluations for specific paper submissions:

• "Software Rasterization of 2 Billion Points in Real-Time"
  In branch compute_rasterizer_2022
  paper - video

• "Rendering Point Clouds with Compute Shaders and Vertex Order Optimization"
  In branch compute_rasterizer_2021

Features and Limitations

• Renders up to one billion points in about 8 milliseconds (hence 2 billion points in real-time, 60fps) on an RTX 3090.
• You need to make sure not to load more than your GPU memory can handle. You'll need about 1.6GB for every 100 million points, plus 1GB or 2GB overhead.
• Drag & Drop a LAS or LAZ files into the window to load it. Only RGB attributes are displayed.
• Requires Windows and NVIDIA GPUs. Pull requests for AMD support are welcome.

Building

• Clone the repository
• Compile build/ComputeRasterizer.sln with Visual Studio 2022.
• Run (ctrl + f5)

Method	Location
basic	./modules/compute_loop_las
prefetch	./modules/compute_loop_las2	fastest, each thread fetches 4 points at a time
hqs	./modules/compute_loop_las_hqs	High-Quality Shading
LOD	./modules/compute_loop_nodes	Support for the Potree LOD format
LOD hqs	./modules/compute_loop_nodes_hqs

Citing

@article{SCHUETZ-2022-PCC,
  title =      "Software Rasterization of 2 Billion Points in Real Time",
  author =     "Markus Sch\"{u}tz and Bernhard Kerbl and Michael Wimmer",
  year =       "2022",
  month =      jul,
  journal =    "Proc. ACM Comput. Graph. Interact. Tech.",
  volume =     "5",
  pages =      "1--16",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2022/SCHUETZ-2022-PCC/",
}

@article{SCHUETZ-2021-PCC,
  title =      "Rendering Point Clouds with Compute Shaders and Vertex Order Optimization",
  author =     "Markus Sch\"{u}tz and Bernhard Kerbl and Michael Wimmer",
  year =       "2021",
  month =      jul,
  doi =        "10.1111/cgf.14345",
  journal =    "Computer Graphics Forum",
  number =     "4",
  volume =     "40",
  pages =      "115--126",
  keywords =   "point-based rendering, compute shader, real-time rendering",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2021/SCHUETZ-2021-PCC/",
}