A high-performance, GPU-driven voxel engine written in C++23 and using OpenGL 4.6, GLM, GLAD.
Table of Contents
Voxel Engine is a high-performance, GPU-driven voxel engine built in C++23 on top of OpenGL 4.6. It's designed around a "GPU does the heavy lifting" philosophy: chunk visibility, meshing requests, and draw submission are all driven through compute shaders and indirect rendering rather than CPU-bound per-frame work, allowing large, dynamic voxel worlds to stream and render with minimal CPU overhead.
At the core of the engine is a binary greedy meshing algorithm that collapses contiguous voxel faces into compressed, GPU-friendly quads, paired with a lock-free VRAM mesh cache and AZDO rendering techniques (persistent mapped buffers, multi-draw indirect, direct state access) to keep draw calls and memory traffic to a minimum. A GPU frustum culling pipeline evaluates chunk visibility every frame, pulling cached meshes straight into the indirect draw buffer while kicking off meshing for anything not yet cached—so newly visible terrain appears with minimal stalling. A 3D ring buffer keeps only the chunks around the camera loaded, streaming in new chunks as the camera crosses boundaries instead of rebuilding the entire world volume.
Beyond rendering, the engine provides a pluggable world generation system via the ChunkGeneratorStrategy interface, real-time voxel editing with undo/redo and brush tools, DDA-based voxel ray-casting for precise picking, and built-in tooling for debugging and performance analysis—including gizmos, an X-macro–driven logging system, and Hazel-inspired JSON trace profiling viewable in chrome://tracing.
-
Compiler: C++23 capable compiler (GCC 11+, Clang 14+, MSVC 2019/2022 with C++23 support).
-
Graphics: GPU and drivers supporting OpenGL 4.6.
-
Libraries: GLM, GLAD, spdlog, stb (or use the bundled loader)
-
Tools: CMake 3.20+, a build tool (make, ninja, or MSBuild).
Clone the repository and set up a build directory:
git clone https://github.com/omar-owis/VoxelEngine.git
cd VoxelEngine && mkdir build && cd buildConfigure the project with CMake:
cmake .. -DCMAKE_BUILD_TYPE=Release [BUILD OPTIONS]The engine exposes several CMake options to customize logging, profiling, and chunk configuration. These can be passed via -D{option}={value} during configuration.
| Option | Type | Default | Allowed Values | Description |
|---|---|---|---|---|
| VE_BUILD_TEST | BOOL |
OFF |
ON, OFF |
Builds the Voxel Engine test suite. |
| VE_LOGGING_ENABLED | BOOL |
ON |
ON, OFF |
Enables engine logging (spdlog). See usage section to configure logging |
| VE_PROFILING_ENABLED | BOOL |
ON |
ON, OFF |
Enables profiling and Google Trace export. see usage section to use profiling |
| VE_CHUNK_TYPE | STRING |
CHUNK8 |
CHUNK8, CHUNK16, CHUNK32 |
Selects the chunk size used by the engine. Adds one of: 8x8x8 Chunk Size, 16x16x16 Chunk Size, 32x32x32 Chunk Size. Invalid values produce a fatal CMake error. |
Note: CHUNK32 is not supported yet. The current compressed‑quad format used in chunk meshes cannot fit the required xyzwh data for each quad into the existing 2‑byte encoding. Work is underway to extend the compression scheme so that CHUNK32 can be fully supported in a future update.
| Input | Action |
|---|---|
| W / A / S / D | Move forward / left / backward / right |
| Space | Move up |
| Left Shift | Move down |
| Mouse Movement | Look around (update camera direction) |
| Mouse Scroll | Adjust camera movement speed |
| Left Mouse Button | Remove targeted voxel |
| Escape | Close the application |
| F1 | Toggle profiling (Google Trace) |
| F2 | Toggle draw mode (Fill ↔ Wireframe) |
| F3 | Toggle chunk bounds visualization |
| F4 | Toggle DebugUI Overlay |
| Tab | Switch between Main Camera and isometric Camera; toggle frustum outline |
- Binary Greedy Meshing: Bitmask-based greedy quad generation that merges contiguous visible voxels per face into large binary-packed quads, drastically reducing triangle count and while preserving exact chunk silhouettes.
-
VRAM Chunk Mesh Cache: Lock-free, gpu lookups, FIFO Page replacement
-
AZDO rendering techinques: Persistent Mapped Buffers, Multi-Draw Indirect, Direct State Access
-
Triple-buffering: indirect command buffer and other rendering buffers split into three rotating regions so the CPU updates one segment while the GPU consumes the other two, eliminating write–read conflicts.
-
Lock-free mutlithreading: Custom thread pool enabling concurrent chunk meshing and lock‑free uploads into the mesh cache,
-
GPU Frustum Culling pipeline: Compute shader evaluates chunk visibility each frame, pushes visible cached meshes into the indirect draw buffer, and issues GPU-side meshing requests for uncached chunks to render on the next frame.
-
Quad compression & Instancing: Quads are packed into a 2‑byte format encoding their local XYZ position, width, height, face direction, and RGBA color
-
Face culling: OpenGL face culling
-
3D ring buffer: Keeps a rotating window of chunks around the camera, loading only the incoming planes when movement crosses a chunk boundary. Avoids regenerating the full volume around the camera, preserves cache locality, and guarantees constant‑time chunk indexing.
- Gizmos: Lightweight debug gizmos for visualizing chunk bounds, ray hits, frustums, and editor handles, using depth‑aware screen‑space rendering.
- Logging: X‑macro–driven logging system backed by spdlog, making it trivial to add new log channels while keeping output fast and thread‑safe.
- Profiling: Lightweight function‑level profiling inspired by Hazel, writing JSON trace files for startup, runtime, and shutdown that can be inspected in chrome://tracing.
- Ray-casting: Voxel ray‑casting uses the Amanatides & Woo DDA grid traversal algorithm for precise, step‑wise intersection across chunk boundaries. Supports accurate voxel picking for editing, selection, and gizmo interaction.
- Voxel editing: Real‑time voxel modification with automatic dirty‑chunk tracking and incremental remeshing. Includes undo/redo support and basic brush tools for painting, carving, and filling.
The ApplicationConfig struct defines all runtime parameters for the engine. These values determine window size, threading behavior, chunk loading distances, cache layout, and the initial world position.
ApplicationConfig appConfig = {
.ScreenWidth = 800,
.ScreenHeight = 600,
.Name = "VoxelEngine",
.ThreadWorkers = std::thread::hardware_concurrency(),
// Cache configuration
.cachePageSize = 400,
.cacheNumOfPages = 62500,
.averageIndirectCmdsPerChunk = 3,
// Chunk distances
.loadedChunkDistance = 15,
.renderChunkDistance = 15,
// Starting world position
.startingWorldPos = glm::vec3(1.0f),
};| Field | Type | Description |
|---|---|---|
| ScreenWidth | uint32_t |
Width of the application window in pixels. |
| ScreenHeight | uint32_t |
Height of the application window in pixels. |
| Name | std::string |
Name of the application; appears in the window title. |
| ThreadWorkers | uint32_t |
Number of worker threads used for meshing & upload tasks (typically std::thread::hardware_concurrency()). |
| cachePageSize | uint32_t |
Number of quads stored per cache page. Controls paging granularity. |
| cacheNumOfPages | uint32_t |
Total number of cache pages available for chunk storage. |
| averageIndirectCmdsPerChunk | uint32_t |
Renderer tuning parameter estimating average indirect draw commands per chunk. |
| loadedChunkDistance | uint32_t |
Length of cubic volumn (in chunks) around the camera where chunks are fully loaded. |
| renderChunkDistance | uint32_t |
Length of cubic volumn (in chunks) where chunks are rendered. Typically equal to loadedChunkDistance. |
| startingWorldPos | glm::vec3 |
Initial world-space position of the camera. |
Logging is configured through a LogConfig struct, allowing control over subsystem verbosity. Each subsystem can be independently set to Trace, Debug, Info, Warn, Error, Critical, or Off.
LogConfig logConfig = {
.CORE = LogLevel::Info,
.RENDERER = LogLevel::Info,
.INPUTS = LogLevel::Info,
.CHUNK = LogLevel::Info,
.VOXEL_MESHER = LogLevel::Info,
.SCENE = LogLevel::Info,
.GPU_BUFFER = LogLevel::Info,
.VOXEL_ENGINE = LogLevel::Info,
};Note: Disabling VE_LOGGING_ENABLED removes all runtime log output. No messages from any subsystem will appear in the console.
Chunk generation is handled through the ChunkGeneratorStrategy interface. Each strategy defines how voxel data is produced for a chunk when the engine requests it. The engine passes the chunk’s coordinates and a writable chunk instance, and the strategy fills its voxels.
A generator implements:
void Generate(const glm::ivec3& chunkCoords,
const std::shared_ptr<ChunkType>& outChunk);Example 3D Perlin Noise:
class Simple3DPerlinNoiseGeneration : public ChunkGeneratorStrategy<ChunkType>
{
public:
Simple3DPerlinNoiseGeneration() : m_PerlinNoise(this->p_Seed) {}
void Generate(const glm::ivec3& chunkCoords,
const std::shared_ptr<ChunkType>& outChunk) override
{
const int randColor = this->p_Dist(this->p_Engine);
for (int x = 0; x < ChunkType::Size; x++)
for (int y = 0; y < ChunkType::Size; y++)
for (int z = 0; z < ChunkType::Size; z++)
{
float n = m_PerlinNoise.Noise3D(
(chunkCoords.x * ChunkType::Size + x) * 0.1f,
(chunkCoords.y * ChunkType::Size + y) * 0.1f,
(chunkCoords.z * ChunkType::Size + z) * 0.1f
);
if (n > 0.2f)
outChunk->SetVoxel(x, y, z, randColor);
}
}
std::string ToString() override { return "Simple3DPerlinNoiseGeneration"; }
private:
PerlinNoise m_PerlinNoise;
};To define custom terrain or structures, inherit from ChunkGeneratorStrategy, override Generate(), and fill the chunk using any logic—noise, heightmaps, patterns, or procedural rules. Then pass the ChunkGenertorStratgy to the engine application and the engine will automatically use your strategy for all chunks.
Refer to chunk_generator_strategy.h for built‑in generation strategies you can use or extend.
The engine includes lightweight function‑level profiling inspired by The Cherno’s Hazel Engine. When profiling is enabled, the engine automatically records timing data into three JSON trace files:
-
Profile-Startup.json: events during engine initialization
-
Profile-Runtime.json: events captured while runtime profiling is active
-
Profile-Shutdown.json: events during engine shutdown
You can open any of these files in a profiling viewer such as Google Trace (chrome://tracing) by dragging the JSON file into the viewer.
Use the F1 key to start or stop runtime profiling. When profiling is disabled via the build option VE_PROFILING_ENABLED=OFF, no profiling files are generated.
Benchmark results were collected using the environment and configuration described below. All tests were executed with logging disabled.
| Component | Specification |
|---|---|
| CPU | Intel i5-10300H @ 2.50 GHz |
| Cores | 8 |
| GPU | NIVIDA GeForce RTX3050 |
| RAM | 16 GB |
| OS | Windows 11 |
| Cache Parameter | Value |
|---|---|
| cachePageSize | 400 quads/page |
| cacheNumOfPages | 62,500 pages |
| averageIndirectCmdsPerChunk | 3 |
The following benchmarks were generated using the Simple3DPerlinNoiseGeneration chunk‑generation strategy, with multi‑threading disabled and both LoadedChunkDistance and RenderChunkDistance set to 31.
| ChunkType | Meshing Time per Chunk | Mesh Upload time |
|---|---|---|
| CHUNK8 | 0.020 ms | 0.050 ms |
| CHUNK16 | 0.064 ms | 0.051 ms |
| CHUNK32 | --- | --- |
MIT License See LICENSE for more information