Draco Compressed Meshes with glTF and 3D Tiles

Khronos glTF Draco compression extension

glTF now has the KHR_draco_mesh_compression extension, which enables loading buffers containing Draco compressed geometry. Starting with Cesium 1.44, we support loading glTF assets with Draco compressed data by leveraging Google’s open-source JavaScript decompression library.

Using compressed meshes reduces the resulting file size of a glTF model, meaning these assets take up less space, download less data, and stream faster. To illustrate the difference, we compressed the following files using the Draco encoder, with the default compression level of 7 for all attributes.

Here’s the glTF 2.0 Draco Compressed Cesium Milk Truck sample model compared to the glTF 2.0 Cesium Milk Truck sample model, both of which contain textures as well as animations.

*Images are sped up at 2x speed for demonstration.

Performance Optimizations

Cesium’s implementation of Draco decoding takes advantage of both asynchronous decoding with Web Assembly and dequantization on the GPU, which means decoding several models (or parts of models) in parallel, as well as less overall memory used.

Parallel Decompression

Cesium takes advantage of Web Workers to decode multiple meshes in parallel. In the case of 3D Tiles, this means multiples tiles can be streamed and decoded simultaneously. Furthermore, each primitive (or part) of a mesh can be decoded separately for faster decoding of complex models. We can retrieve each segment of the encoded buffer and pass the data to separate workers to asynchronously decode in parallel before returning the data necessary to render the mesh to the main thread.

When supported by the browser, we load and compile the decoding module Web Assembly binary and share it across multiple workers, further increasing the speed of decompression as opposed to using a pure JavaScript solution.

Dequantization on the GPU

Cesium also decodes some attributes on the GPU, decoding outside of the main thread and using less memory. Vertex attributes that are usually stored as 32-bit floating point numbers, such as position attribute data, can be decoded as quantized 16-bit integer values. Additionally for attributes that are unit vectors such as normals, we can decode as oct-encoded data.

When decoding on the GPU, we skip the quantization or octahedron transform operation in the Draco decoder module and instead retrieve and store any transformation constants. The smaller decoded data can be passed to the GPU where the dequantization or oct-decoding operations are performed in the shader when rendering. This results in a smaller memory footprint, both in the main application thread running on the CPU and in parallel on the GPU.

When performing dequantization in the GPU with the previously mentioned New York City 3D tileset, there is a 52% savings in memory used by the GPU with no difference in file size, no differences in visual quality, and no impact on total tileset loading time as compared to decoding entirely with the Draco decoding module.