Real-Time Hybrid Hair Renderer in Vulkan™

This is a proof of concept hair renderer that's based on a novel hybrid approach. It's capable of rendering strand-based hair geometry in real-time, even for fully simulated hair styles with over 100,000 strands. Our hybrid rendering pipeline scales in the performance/quality domain by using a rasterizer for closeup shots and a raymarcher for level of detail minification. We can render multiple hair styles at once with this hybrid technique, which has a smooth transition between the two solutions.

In the figure above we see the ponytail from TressFX 3.1 (with 136,320 hair strands and 1,635,840 line segments) rendered in real-time (7ms) with our rasterized solution. If we want to render more than a couple of characters on the screen at once, we're going to need a more scalable solution, as it doesn't scale for far away distances. This is where our raymarcher comes in, as its performance scales linearly with the fragments rendered on the screen, and is a also lot cheaper for far away hairs. However, our raymarcher's performance breaks down during close-up shots, and also looks "worse" than our rasterizer for those cases, as it's an approximation of the real geometry. The trick is to use both of the solutions together! The rasterizer may perform worse, but it still produces high-quality close-up shots, and performance scaling isn't too bad in those cases. Our raymarcher performs and scales better at a distance and is indistinguishable from the rasterized results in those cases. The figure below shows both the rasterized and raymarched solutions together. Each screenshot is split in the middle, with the left part being the rasterized solution, and the right side the raymarched solution. We have alpha blended these two in the middle to simulate a level of detail transition. As you can see the results are quite close and the transitions are not that noticeable from far away. There's however a huge (up to 5-6x) performance difference that scales proportional to distance.

It is written 100% from scratch in Vulkan™ and simulates light scattering, self-shadowing and transparency in real-time. We do anti-aliasing by using a fast line coverage method. Our volumetric approximation is derived from the original geometry by a fast voxelization scheme that's used for raymarching and direct shading on an isosurface. We also use this volume for ambient occlusion and for more precise self-shadowing even in the rasterized case. Our work was published at EGSR 2019.

Outline

For the rest of this document you'll find out about the features in our renderer and the benefits of using a hybrid approach when rendering hair, by looking at its performance. Once you're intrigued, we'll show you how to build, and run the project yourself so you can try it out. I have written lots of documentation if you want to find out how it works or which limitations there are. You'll find more screenshots towards the end of this document and in the Captain's Log found in the project wiki.

Introduction
Outline
Table of Contents
Features
Benchmarks
Dependencies
Compiling
System Requirements
Usage
Documentation
Directories
Reporting Bugs
Acknowledgements
Legal Notice
Screenshots

Features

A real-time hybrid hair rendering pipeline suitable for video games, that scales in the performance and quality domain. It is:

Written from scratch in modern C++17 with minimal dependencies,
Uses the Vulkan™ API with a lightweight wrapper: vkpp, written for modern C++17, with proper lifetime management,
Has a built-in raytracer based on Intel's Embree® with a CMJ sampler to compare ground-truth global effects, like AO,
Loads Cem Yuksel's free & open .hair file format, and has a easy human-readable scene graph format based on JSON,
Consists of a strand-based hair rasterizer and a volume raymarcher.

It uses this rasterized solution for close-up shots, and our raymarched solution for level of detail. This hybrid hair renderer:

Models single light scattering in a strand with Kajiya-Kay's shading,
Estimates hair self-shadowing with a fast Approximated Deep Shadow Map (ADSM) method à la Tomb Raider (2013),
Produces anti-aliased strands by using a simple, but effective, line coverage calculation similar to Emil Persson's GPAA,
Resolves strand transparency with a fragment k-Buffer PPLL similar to TressFX's OIT that's built and sorted on the GPU,
Has a scalable level of detail scheme based on volume ray casting.

This novel volumetric approximation for strand-based hair can be found once per-frame for fully simulated hair. It features:

A very fast compute-based strand voxelization technique for hairs,
An approximation of Kajiya-Kay's model by finding the tangents inside of a volume by quantized strand voxelization,
An ADSM equivalent, that also takes into account the varying hair spacing by using the actual strand density as input,
A way to approximate the local ambient occlusion by using the same strand density (a useful representation for hair),
An approximation of transparency (no DVR!) for low density areas.

Our hybrid rendering solution combines the best of strand- and volume-based hair representations. Some benefits are that:

It is faster than purely raster-based techniques in the far away case,
The performance is more predictable and configurable as raymarching scales linearly with the hair's screen coverage,
The level of detail transition is quite smooth because both the rasterizer and raymarcher approximate similar effects,
The ambient occlusion and other global effects are trivial to estimate in a volume but impossible in a pure rasterizer,
It is automatic as our voxelization works even with simulated hairs.

Benchmarks

Along with this project we bundle a set of benchmarks that can be run by passing the --benchmark yes flag. They compare the performance between the rasterizer and raymarcher and how these performance scale (e.g. with respect to increasing distances or strands). In order for you to get an idea if our solution is good enough for your purposes, we have included the results from our paper, which were run on a Radeon™ Pro WX 9100. The results were taken with V-Sync off and without any other GPU intensive programs running in the background. The timing information was taken via Vulkan timestamp queries, and averaged over a period of 60 frames (not much variance). We have plotted the results below for your viewing pleasure.

In the above plot to the left we see how the rasterizer and raymarcher fare at different distances, and how much time each rendering pass takes. For the near cases (e.g. a character close-up) the raymarched solution is around twice as fast, but the fidelity isn't as good, as strands appear to be clumped together because of the volume approximation. On the other hand, the rasterized solution produces high-quality output as each strand is individually distinguishable. However for far away to medium distances, these small details are not noticable, and the rasterized and raymarched solution are indistinguishable. The raymarcher on the other hand is now 5x faster in these distances! It has better scaling with distance for far away shots.

Setup: Rendering at 1280x720, Ponytail Scene, V-Sync Off, 1024x1024 Shadow Maps, 256³ Volume, 512 Raymarching Steps.

The raymarcher also doesn't have to produce shadow maps, which would scale linearly with the number of light sources for the scene. Finally, notice that the strand voxelization is quite cheap and does not account for much of the total render time. In the memory department, the figure to the right shows the GPU data breakdown. When comparing to the original strand-based geometry, the volume does not consume an inordinate amount of memory, and this value can also be tweaked with the volume resolution. The main culpit are the PPLL nodes that are used for our transparency solution. These scale with the resolution and also depend on how many strands are being shaded (and might lead to artifacts if memory underallocated).

For the two plots above we see how performance scales for each renderer with respect to screen coverage and number of hair strands. The raymarcher has a lower intercept, making it cheap to render for low screen coverage (far away distances). Performance on the rasterizer scales linearly with the number of hair strands (as expected), and also for the raymarcher but with a very slow slope (caused by the voxelization). Our technique works especially well for realistic amounts of hair, where anything less than ~20,000 strands of hair will look bald. While the scaling on the right doesn't look very promising for the raymarcher, its performance can be tuned by changing the number of raymarch steps that moves the intercept up / down.

Dependencies

premake5 (pre-build)
Any Vulkan™ 1.1 SDK
glfw3 (tested v3.2.1)
embree3 (uses v3.2.4)
Any C++17 compiler!

All other dependencies are fetched using git submodules. They include the following awesome libraries: g-truc/glm, ocurnut/imgui, syoyo/tinyobjloader, nothings/stb and nlohmann/json. The C++17 Vulkan wrapper: vkpp is being developed alongside this project. It will at a later time be split into another repository: vkpp, when I have time to do it.

Compiling

First, make sure you've changed your current directory to vkhr
Then do git submodule update --init --recursive --depth 1
- Description: fetches submodule dependencies to foreign
Since we use premake, you'll most likely need to fetch it as well:
- Tip: there's pre-generated Visual Studio solutions in build
  - if you're happy with that, you can skip the steps below
- Unix-like: just install premake5 with your package manager
Now make sure you have the glfw3 external dependency solved
- Unix-like: just install glfw with your package manager too
- Visual Studio: pre-built version is already provided for you!
Finally, you'll also need Embree for the hair raytracing back-end:
- Unix-like: just install embree using your package managers
- Visual Studio: pre-built version is already provided for you!
Generate the vkhr project files by targeting your current setup
- Visual Studio: premake5 vs2019 or my alias make solution
  - then open the Visual Studio project in build/vkhr.sln
  - you might have to retarget the VS solution to your SDK
- GNU Makefiles: premake5 gmake or just call make all/run.
Build as usual in your platform, and run with bin/vkhr <scene>.

Distribution

Install: if you're on Arch Linux it's as simple as running makepkg -i.

For Windows just call make distribute for a "portable" ZIP archive.

The client then only needs a working Vulkan runtime to start vkhr.

System Requirements

Platforms must support Vulkan™.

It has been tested on these GPUs:

NVIDIA® GeForce® MX150,
Radeon™ Pro WX 9100 Graphics,
Intel® HD Graphics 620.

on Windows 10 and GNU / Linux.

Usage

bin/vkhr: loads the default vkhr scene share/scenes/ponytail.vkhr with the default render settings.
bin/vkhr <settings> <path-to-scene>: loads the specified vkhr scene, with the given render settings.
bin/vkhr --benchmark yes: runs the default benchmark and saves it to a CSV file inside benchmarks/.
- Plots can be generated from this data by using the utils/plotte.r script (requires R and ggplot).
Default configuration: --width 1280 --height 720 --fullscreen no --vsync on --benchmark no --ui yes
Shortcuts: U toggles the UI, S takes a screenshots, T switches between renderers, L toggles light rotation on/off, R recompiles the shaders by using glslc (needs to be set in $PATH to work), and Q / ESC quits the app.
Controls: simply click and drag to rotate the camera, scroll to zoom, use the middle mouse button to pan.
UI: all configuration happens in the ImGUI window that is documented under the Help button in the UI.
man docs/vkhr.1 will open the manual page containing even more detailed usage information for vkhr.

Documentation

You're reading part of it! Besides this readme.md, you'll find that most of the important shaders are nicely documented. Two good examples are GPAA.glsl for the line coverage calculations, and approximate_deep_shadows.glsl for the self-shadowing technique. You'll notice that the quality of it varies quite a bit, feel free to open an issue if you sense something isn't clear. I haven't documented the host-side of the implementation yet as that would take too long, and isn't that interesting anyway.

If you want a high-level summary of our technique read Real-Time Hybrid Hair Rendering, which is a short conference paper on our method (only the pre-print). You'll also find a copy of it here, which you can build by using LaTeX. If you want a more extensive and detailed version of our paper, my thesis Scalable Strand-Based Hair Rendering, will soon be available. Both of these also show the difference between our technique and other existing frameworks like TressFX, that only use a rasterizer.

And if you still haven't had enough, I have written a bunch of entries in the Captain's Log, that shows the progress log from day 1 to the current version. Besides having a lot of pretty pictures, it shows the problems we encountered, and how we've solved them. This gives a bit more insight into why we have chosen this approach, and not something completely different. The slides for my thesis defense and presentation at EGSR 2019 could also be useful to get an overview into our technique.

Directories

benchmarks: output from the bundled benchmarks goes in here.
bin: contains the built software and any other accompanying tools.
build: stores intermediate object files and generated GNU Make files.
- obj: has all of the generated object files given under compilation.
- Makefile: automatically generated by executing premake5 gmake.
- *.make: program specific make config for augmenting Makefile.
- you'll also find the pre-generated Visual Studio '17 solution here.
docs: any generated documentation for this project is over here.
foreign: external headers and source for libraries and modules.
include: only internal headers from this project should go here.
- vkhr: internal headers for the Vulkan hair renderer project.
- vkpp: headers for a minimal modern C++ Vulkan wrapper.
license.md: please look through this very carefully.
premake5.lua: configuration file for the build system.
readme.md: this file contains information on the project.
share: any extra data that needs to be bundled should go here.
- images: any images on disk that should be used as textures.
- models: the meshes/models/materials to be used in the project.
- shaders: all of the uncompiled shaders should go over here.
- scenes: any sort of scene files (e.g. in json) should go here.
- styles: the hair styles compatible with the Cem Yuksel format.
src: all source code for the project should be located below here.
- vkhr: source code for the Vulkan hair renderer project itself.
- vkpp: full implementation of an Vulkan C++ wrapper (separate).
- main.cc: the primary entry point when generating the binaries.
utils: any sort of helper scripts or similar should be over here.

Reporting Bugs

vkhr is 100% bug-free, anything that seems like a bug is in fact a feature!

This is a proof-of-concept research prototype, and as such, I wouldn't recommend using it for something serious, at least as it is. Also, do not expect this repository to be well maintained, I will not spend too much time with it after the thesis is done.

Still, if you find anything, feel free to open an issue, I'll see what I can do :)

Acknowledgements

First I would like to thank Matthäus Chajdas, Dominik Baumeister, and Jason Lacroix at AMD for supervising this thesis, and for always guiding me in the right direction. I'd also like to thank the fine folk at LiU for providing feedback and support, in particular, my examinator Ingemar Ragnemalm and Harald Nautsch at ISY and Stefan Gustavson from ITN. I would also like to thank AMD and RTG Game Engineering for their hospitality and friendliness, and for letting me sit in their Munich office.

Legal Notice

Vulkan and the Vulkan logo are registered trademarks of Khronos Group Inc.

All hair styles are courtesy of Cem Yuksel's great HAIR model files repository.

The ponytail and bear hair geometry are from the TressFX 3.1 repository, and proper rights have been granted by AMD Inc. to be used in this repository. However, you are not allowed to use it outside of this repository! i.e. not the MIT license for it!

The woman model was created by Murat Afshar (also for Cem Yuksel's repo).

Everything in this repository is under the MIT license except the assets I've used. Those fall under the license terms of their respective creators. All of the code in this repository is my own, and that you can use however you like (under the license).

Both GLFW and Embree are pre-compiled to facilitate building on Windows.

See: foreign/glfw/COPYING and foreign/embree/LICENSE for these licenses.