A multi-platform OpenGL / Vulkan based game engine that works on Desktop Java and browsers with WebGL2.
I started working on a graphical scene editor.
The editor is still in an early state and not very useful yet, but you can try the Web Demo
in case you are curious. However, there are a few rather essential features, which are only supported by the JVM
version of the editor (e.g. hot-reloading code and importing additional assets). You can launch the JVM editor by
cloning this project and running the gradle task :kool-editor-template:runEditor from a command line or your IDE).
Apart from that, this is very much a code-only game engine. I recommend taking a look at the demos listed below in case
you are curious (all demo source code is available in the kool-demo subproject). The library is also published on
maven central, see below.
I also made an actual game with this: Blocks and Belts. Give it a try (it's free)!
Feel free to join the Discord Server!
- Editor: Work in progress graphical editor for this engine. Fully implemented within the engine itself.
- Island: Height-map based island incl. some wind-affected vegetation + a basic controllable character.
- Physics - Ragdoll: Ragdoll physics demo.
- Physics - Vehicle: A drivable vehicle (W, A, S, D / cursor keys, R to reset) based on the Nvidia PhysX vehicles SDK.
- Physics - Joints: Physics demo consisting of a chain running over two gears. Uses a lot of multi shapes and revolute joints.
- Physics - Collision: The obligatory collision physics demo with various different shapes.
- Embedded UI: Integrated UI framework implemented completely within the engine. Fast, highly customizable and easy-to-use.
- CPU Particles: Two teams of bees fighting against each other. Purely CPU based (single-threaded even) so not as fast as it could be(e). Also, JVM is a lot faster than JS.
- Fluffy Bunny: Shell-shading based rendering of animated fur (based on this video).
- Creative Coding: A few relatively simple demos showcasing different techniques of generating procedural geometry.
- Procedural Geometry: Small test-case for procedural geometry; all geometry is generated in code (even the roses! Textures are regular images though). Also, some glass shading (shaft of the wine glass, the wine itself looks quite odd when shaded with refractions and is therefore opaque).
- glTF Models: Various demo models loaded from glTF / glb format
- Flight Helmet from glTF sample models
- Polly from Blender
- Coffee Cart from 3D Model Haven
- Camera Model also from 3D Model Haven
- A few feature test models also from the glTF sample model repository
- Deferred Shading: Thousands of dynamic light sources, bloom and ambient occlusion.
- Screen-space Ambient Occlusion: Roughly based on this article by John Chapman with slightly optimized sampling (also shamelessly recreated his demo scene).
- Screen-space Reflections: A simple PBR shaded model with screen-space reflections and up to four spot-lights with dynamic shadows.
- Physical Based Rendering: Interactive PBR demo with image based lighting for various materials and environments (underlying PBR theory from this awesome article series).
- Instanced / LOD Drawing: Instanced rendering demo of the Stanford Bunny. Uses six levels of detail to render up to 8000 instances.
- Mesh Simplification: Interactive mesh simplification demo (based on traditional error-quadrics)
Code for all demos is available in kool-demo sub-project.
If you are adventurous, you can use kool as a library in your own (multiplatform-) projects. The library is published on maven central, and there is a separate repo containing a minimal template project to get you started:
https://github.com/fabmax/kool-templates
The demos mentioned above and examples shown below should give you a rough idea on how to do stuff (documentation is still a bit of a weak spot).
You can launch the desktop demo app directly from a terminal via gradle with ./gradlew :kool-demo:run.
Running the main() method from within IntelliJ requires that the native libraries are located in a local folder and added as file dependencies (seems to be some kind of dependency resolution bug in IntelliJ when importing multiplatform projects with JVM runtimeOnly libraries).
The required libs are copied automatically on build. So, in order to launch the demos from within IntelliJ you need
to build the project first (or manually run the cacheRuntimeLibs task) and then re-sync the gradle project, so that
the libs are resolved and added to the IntelliJ module classpath.
- Basic Compute shader support (OpenGL only for now, only storage textures)
- Reversed-depth rendering for vastly improved depth precision and range (more or less infinite)
- Physics simulation (based on Nvidia PhysX 5.1, using physx-jni on Java and physx-js-webidl on javascript)
- Kotlin DSL based shader language (translates into GLSL)
- Neat little integrated GUI framework. The API is heavily inspired by Jetpack Compose but the implementation is different, as it needs to run within the OpenGL context.
- MSDF Font support for text rendering in arbitrary font sizes
- Experimental Vulkan rendering backend (on JVM)
- Support for physical based rendering (with metallic workflow) and image-based lighting
- (Almost) complete support for glTF 2.0 model format (including animations, morph targets and skins)
- Skin / armature mesh animation (vertex shader based)
- Deferred shading
- HDR lighting with Uncharted2 tone-mapping
- Optional Screen-space reflections
- Optional Bloom
- Screen-space ambient occlusion
- Normal, roughness, metallic, ambient occlusion and displacement mapping
- Lighting with multiple point, spot and directional lights
- Shadow mapping for multiple light sources (only spot and directional lights for now)
- Basic audio support
Getting a basic scene on the screen is quite simple:
fun main() = KoolApplication { ctx ->
ctx.scenes += scene {
defaultOrbitCamera()
addColorMesh {
generate {
cube {
colored()
}
}
shader = KslPbrShader {
color { vertexColor() }
metallic(0f)
roughness(0.25f)
}
onUpdate {
transform.rotate(45f.deg * Time.deltaT, Vec3f.X_AXIS)
}
}
lighting.singleDirectionalLight {
setup(Vec3f(-1f, -1f, -1f))
setColor(Color.WHITE, 5f)
}
}
}The above example creates an application with a single scene and sets up a mouse-controlled camera
(with defaultOrbitCamera()).
As you might have guessed the addColorMesh { ... } block creates a colored cube and adds it to the scene.
In order to draw the mesh on the screen it needs a shader, which is assigned with
shader = KslPbrShader { ... }. This creates a simple PBR shader for a dielectric material
with a rather smooth surface. Color information is taken from the corresponding vertex attribute.
The onUpdate-block is called on each frame and modifies the cube transform to rotate it 45° per second around its
X-axis.
Finally, we set up a single directional scene light (of white color and an intensity of 5), so that our cube can shine
in its full glory. The resulting scene looks like this.
Model loading, animation and more advanced lighting with shadow mapping and ambient occlusion requires only a few more lines of code:
fun main() = KoolApplication { ctx ->
ctx.scenes += scene {
defaultOrbitCamera()
// Light setup
lighting.singleSpotLight {
setup(Vec3f(5f, 6.25f, 7.5f), Vec3f(-1f, -1.25f, -1.5f), 45f.deg)
setColor(Color.WHITE, 300f)
}
val shadowMap = SimpleShadowMap(this, lighting.lights[0])
val aoPipeline = AoPipeline.createForward(this)
// Add a ground plane
addColorMesh {
generate {
grid { }
}
shader = KslPbrShader {
color { constColor(Color.WHITE) }
shadow { addShadowMap(shadowMap) }
enableSsao(aoPipeline.aoMap)
}
}
// Load a glTF 2.0 model
launchOnMainThread {
val materialCfg = GltfFile.ModelMaterialConfig(
shadowMaps = listOf(shadowMap),
scrSpcAmbientOcclusionMap = aoPipeline.aoMap
)
val modelCfg = GltfFile.ModelGenerateConfig(materialConfig = materialCfg)
val model = Assets.loadGltfModel("path/to/model.glb", modelCfg)
model.transform.translate(0f, 0.5f, 0f)
if (model.animations.isNotEmpty()) {
model.enableAnimation(0)
model.onUpdate {
model.applyAnimation(Time.deltaT)
}
}
// Add loaded model to scene
addNode(model)
}
}
}First we set up the lighting. This is very similar to the previous example but this time we use a spot-light, which requires a position, direction and opening angle. Other than directional lights, point and spot-lights have a distinct (point-) position and objects are affected less by them, the farther they are away. This usually results in a much higher required light intensity: Here we use an intensity of 300.
Next we create a SimpleShadowMap which computes the shadows cast by the light source we defined before.
Moreover, the created AoPipeline computes an ambient occlusion map, which is later used by the shaders to
further improve the visual appearance of the scene.
After light setup we can add objects to our scene. First we generate a grid mesh as ground plane. Default size and
position of the generated grid are fine, therefore grid { } does not need any more configuration. Similar to the
color cube from the previous example, the ground plane uses a PBR shader. However, this time we tell the shader to
use the ambient occlusion and shadow maps we created before. Moreover, the shader should not use the vertex color
attribute, but a simple pre-defined color (white in this case).
Finally, we want to load a glTF 2.0 model. Resources are loaded via the Assets object. Since resource loading is a
potentially long-running operation we do that from within a coroutine launched with launchOnMainThread { ... }.
By default, the built-in glTF parser creates shaders for all models it loads. The
created shaders can be customized via a provided material configuration, which we use to pass the shadow and
ambient occlusion maps we created during light setup. After we created the custom model / material configuration
we can load the model with Assets.loadGltfModel("path/to/model.glb", modelCfg). This suspending function returns the
loaded model, which can then be customized and inserted into the scene. Here we move the model 0.5 units along the
y-axis (up). If the model contains any animations, these can be easily activated. This example checks whether there
are any animations and if so activates the first one. The model.onUpdate { } block is executed on every frame and
updates the enabled animation. The model is inserted into the scene with addNode(model). Calling addNode(model)
from within the coroutine is fine, since the coroutine is launched via launchOnMainThread { ... } and therefor
is executed by the main render thread. If a different coroutine context / thread were used, we had to be careful to
not modify the scene content while it is rendered.
The resulting scene looks like this. Here, the Animated Box from the glTF sample repository is loaded.
Kool comes with an embedded UI framework, which is heavily inspired by Jetpack Compose but was implemented from scratch. Here is a small example:
fun main() = KoolApplication { ctx ->
ctx.scenes += UiScene(clearScreen = true) {
Panel(colors = Colors.singleColorLight(MdColor.LIGHT_GREEN)) {
modifier
.size(400.dp, 300.dp)
.align(AlignmentX.Center, AlignmentY.Center)
.background(RoundRectBackground(colors.background, 16.dp))
var clickCount by remember(0)
Button("Click me!") {
modifier
.alignX(AlignmentX.Center)
.margin(sizes.largeGap * 4f)
.padding(horizontal = sizes.largeGap, vertical = sizes.gap)
.font(sizes.largeText)
.onClick { clickCount++ }
}
Text("Button clicked $clickCount times") {
modifier
.alignX(AlignmentX.Center)
}
}
}
}Here, we create a new UiScene and add a Panel to it, which serves as top-level container for our UI content. Within
the Panel-block, we add a button and a text field. All appearance and layout-properties of the UI elements are
controlled by their modifiers.
Whenever the button is clicked we increment a clickCount which is then displayed by the text field. This works
because the Panel-block is executed each time any remembered state (or mutableStateOf()) within the block changes.
The resulting scene looks like this.
More complex layouts can be created by nesting Row { } and Column { } objects. The
full UI demo should give you an impression on what's possible.
Kool comes with its own shader language (called ksl), which is implemented as a Kotlin Type-safe builder / DSL. The ksl shader code you write is used to generate the actual GLSL shader code. The benefit with this approach is that there is no hard-coded GLSL code in common code, and it should be relatively easy to add different generators which generate shader code for different backends in the future (e.g. WGSL, or metal).
Writing shaders in ksl is very similar to GLSL, here's how a hello-world style shader looks like:
fun main() = KoolApplication { ctx ->
ctx.scenes += scene {
defaultOrbitCamera()
addColorMesh {
generate {
cube {
colored()
}
}
shader = KslShader("Hello world shader") {
val interStageColor = interStageFloat4()
vertexStage {
main {
val mvp = mvpMatrix()
val localPosition = float3Var(vertexAttribFloat3(Attribute.POSITIONS))
outPosition set mvp.matrix * float4Value(localPosition, 1f.const)
interStageColor.input set vertexAttribFloat4(Attribute.COLORS)
}
}
fragmentStage {
main {
colorOutput(interStageColor.output)
}
}
}
}
}
}The interesting part starts at shader = KslShader() = { ... }. Here a new shader is created and assigned to the mesh
created before. If you ever wrote a shader before the structure should be familiar: The shader consists of a vertex
stage (responsible for projecting the individual mesh vertices onto the screen) and a fragment stage (responsible
for computing the output-color for each pixel covered by the mesh). This example shader is almost as simple as a valid
shader can be: It uses a pre-multiplied MVP matrix to project the vertex position attribute to the screen. Moreover,
the color attribute is taken from the vertex input and forwarded to the fragment shader via interStageColor. The
fragment stage then simply takes the color from interStageColor and writes it to the screen.
A little more complex example is available in HelloKsl, which looks like this. Of course, shaders can get more complex than that, you can dig further into the code. All shaders currently used in kool are written in ksl.
After playing around with various different engines on javascript and JVM I came to the conclusion that all of them had some kind of flaw. So I decided to write my own bindings for Nvidia PhysX: physx-jni for JVM, and physx-js-webidl for javascript.
This was quite a bit of work (and is an ongoing project), but I think it was worth it: By writing my own bindings I get the features I need, and, even better, I get the same features for javascript and JVM, which makes the multiplatform approach much easier.