Accelerated Volume Renderer

This is an open-source implementation of Accelerated Volume Rendering with Chebyshev Distance Maps. Please consider citing this work if you use it:

@inproceedings{DeakinAcceleratedvolumerendering2019,
  address = {{Brisbane, QLD, Australia}},
  series = {{{SA}} '19},
  title = {Accelerated Volume Rendering with {{Chebyshev}} Distance Maps},
  isbn = {978-1-4503-6945-9},
  booktitle = {{{SIGGRAPH Asia}} 2019 {{Technical Briefs}}},
  publisher = {{ACM}},
  url = {http://doi.acm.org/10.1145/3355088.3365164},
  doi = {10.1145/3355088.3365164},
  author = {Deakin, Lachlan and Knackstedt, Mark},
  year = {2019},
  keywords = {volume rendering,empty space skipping,distance map,ray casting},
  pages = {25--28},
}

Features

Ray casting volume renderer with efficient empty space skipping
- Gradient map precomputed (compute shader)
- Simple sliders manipulate 2D Transfer Function (TF) texture
- Occupancy map update on TF change used for empty space skipping (compute shader)
- Occupancy map to distance map for faster ray casting (comptue shader)
The viewpoint may enter the volume
- The volume is clipped at some distance from the camera and the vertices of the box-plane intersection are computed in a vertex shader
Volumes are clipped by the depth buffer
Runs in a single subpass with two draw calls per volume

Dependencies

Vulkan-Samples
Some assets from Vulkan-Samples-Assets
Boost

Building

git clone --recurse-submodules <this repository>
You need boost to assist with IO (for now). I recommend grabbing it through vcpkg or just use the system package on linux
The first release includes "assets.zip" which includes the minimum required assets and some volumetric images.
Put any other volumetric images in the newly copied assets folder.
Do a regular cmake build (perhaps using the toolchain file from vcpkg)

Running

  vrender.exe <binary_volumetric_image>...

Make sure the assets and shaders folders are in the working directory. Volumetric images need to be in the assets folder. Multiple volumetric images can be loaded by supplying multiple arguments. The stag beetle image will be loaded if no arguments are supplied.

Volumetric image format

Volumetric images must be densely packed binary files with no header and any extension. For a volume file volume.xyz there should be an accompanying header file file.xyz.header structured like so

1024 1024 795 # extents
0.0003 0.0003 0.0007 # voxel size
0.0 255.0 # normalisation range
uint8_t big # data type and endianness (big or little)
1 0 0 90 # rotation axis and angle (degrees)

This is not a standard format and is just a placeholder

Volumetric images are always converted to uint8_t and normalised from 0-255 based on the normalisation range specified in the header.

Parameters

This example has a number of parameters which can be modified at runtime

ESS: choose between the two empty space skipping (ESS) methods described in the paper (or none), useful to compare performance
ERT: early ray termination, rays terminate when an opacity limit is reached
Sampling: controls the density of voxel samples along rays
Clip distance: controls the distance from the camera where volumes are clipped by a plane
Alpha: all voxel opacities are multiplied by this number
Intensity: Intensity range which maps alpha to [0-1]
Gradient: Gradient range which maps alpha to [0-1] (multiplied with intensity alpha)
Test: output the entry/exit coordinates for the rays or the number of the combined number of texture samples of the volume and distance map ** try changing the empty space skipping method or early ray termination and see how this changes

License

See LICENSE. This example uses Vulkan-samples which has many third-party dependencies, each of which may have independent licensing.

TODO

The IO is incredibly barebones now. needs error checking/file picking/loading and support some common formats for volumetric images
UI for proper texture-based 2D transfer functions
Modifications to vulkan-samples required to use a depth image as both an input attachment and depth stencil attachment simultaneously. Current behaviour:
- If sponza scene hidden, volumes write to depth image
- If sponza scene visible, volumes read from depth image and are clipped accordingly
~~Sparse partially resident textures to support massive sparse images~~.
- Requires too much modification to Vulkan-samples
Pull requests are welcomed

cjsb / VkVolume