Kerntopia is a comprehensive SLANG-centric GPU kernel benchmarking suite that makes kernel development accessible to everyone.
Whether you're a curious developer wanting to demystify SLANG or a SIGGRAPH regular- this suite has your back.
- 🚀 Minimizes the Boilerplate Barrier: Write SLANG kernels, run on CUDA + Vulkan easily
- 🔍 System Interrogation: A hardware detective - show available devices and backend runtimes in use
- 🎯 Three Ways to Run Kernels: Suite mode, standalone executables, or Python wrapper (WIP)
- 📊 Complete Audit Trails: Every benchmark is reproducible with checksums & timestamps (TODO)
- 🎓 Discovery Focused: Heavily commented examples and abstraction to keep the focus your kernel of interest.
Transform this craft scene with a Gaussian blur kernel:
| Before (Original) | After (Gaussian Blur) |
|---|---|
 |
 |
Photo by Giulia Bertelli on StockSnap
# One command, multiple backends, infinite possibilities
❯ ./kerntopia run conv2d
Kerntopia v0.1.0 - SLANG-Centric Kernel Execution Suite
Explore compute kernels with abstracted backend selection, sandboxing, benchmarking, and more.
Running Kerntopia tests with configuration:
Backend: ALL AVAILABLE BACKENDS
Mode: Functional
Compilation: precompiled
Available backends: CUDA, VULKAN
Running tests with filter: *Conv2D*
Note: Google Test filter = *Conv2D*
[==========] Running 2 tests from 1 test suite.
[----------] Global test environment set-up.
[----------] 2 tests from Conv2D_AllBackends/Conv2DParameterizedTest
[ RUN ] Conv2D_AllBackends/Conv2DParameterizedTest.GaussianFilter/CUDA_CUDA_SM_7_0_D0
[ OK ] Conv2D_AllBackends/Conv2DParameterizedTest.GaussianFilter/CUDA_CUDA_SM_7_0_D0 (215 ms)
[ RUN ] Conv2D_AllBackends/Conv2DParameterizedTest.GaussianFilter/VULKAN_GLSL_450_D0
[ OK ] Conv2D_AllBackends/Conv2DParameterizedTest.GaussianFilter/VULKAN_GLSL_450_D0 (85 ms)
[----------] 2 tests from Conv2D_AllBackends/Conv2DParameterizedTest (300 ms total)
[----------] Global test environment tear-down
[==========] 2 tests from 1 test suite ran. (300 ms total)
[ PASSED ] 2 tests.
✅ All 2 tests passed!v0.1.0 - Initial Proof of concept pre-release!
Tested: WSL2 + Ubuntu 24.04 (Win11 host)
Expected: Ubuntu 22.04+ Native
Planned: Windows 11 Native
- CPU: x86_64
- GPU: NVIDIA GTX 1650+ (SM 7.0+) or Intel integrated graphics
- Vulkan: Any Vulkan 1.0+ compatible device (Including CPU emulator)
# Ubuntu 24.04 essentials
sudo apt update
sudo apt install build-essential cmake git
# GPU runtimes (choose your path)
# CUDA: https://developer.nvidia.com/cuda-downloads
# Vulkan SDK: https://vulkan.lunarg.com/sdk/home#linuxCompiler: GCC 13.3.0+ (Ubuntu 24.04), C++17 mode
Ensure you have git-lfs installed (Used for test inputs):
sudo apt-get install git-lfsClone, build, and run:
git clone https://github.com/[username]/kerntopia.git
cd kerntopia && mkdir build && cd build
cmake .. && make -j$(nproc)
# Your GPU benchmark suite is ready!
./kerntopia info --verbose # See what GPUs you have
./kerntopia run conv2d # Run your first kernel# Suite mode - the full GTest experience
./kerntopia run all # Everything, all backends
./kerntopia run conv2d --backend cuda # Specific kernel + backend
./kerntopia run conv2d --backend vulkan --device 0
# Standalone mode - focused kernel program
./kerntopia-conv2d # Quick functional test
./kerntopia-conv2d --backend vulkan --device 0 # Target specific hardware
# System interrogation - know your hardware
./kerntopia info --verbose # Full system analysisflowchart TD
A["`**User Interface Layer**
• kerntopia CLI
• kerntopia-* standalone
• Python wrapper`"] --> B["`**Kernel Management**
• SLANG source kernels
• Precompiled IR cache
• JIT compilation`"]
B --> C["`**Backend Abstraction**
• Unified device interface
• Dynamic runtime loading
• Cross-platform compatibility`"]
C --> D["`**CUDA Backend**
• NVIDIA GPU support
• PTX compilation
• cuDNN integration`"]
C --> E["`**Vulkan Backend**
• Universal GPU support
• SPIR-V compilation
• Cross-vendor compatibility`"]
C --> F["`**Future Backends**
• DirectX 12
• CPU (x86_64)
• Metal (macOS)`"]
D --> G["`**Hardware Layer**
• NVIDIA RTX/GTX
• System interrogation
• Performance monitoring`"]
E --> H["`**Hardware Layer**
• Intel integrated
• AMD discrete
• Multi-vendor support`"]
The architecture eliminates heterogeneous boilerplate - write once in SLANG, run everywhere.
- CUDA (12.0+ SDK) ✅ Tested: NVIDIA RTX 4060 CUDA 13.0 SDK
- Vulkan (1.3) ✅ Tested: Vulkan CPU backend
Coming Soon:
- CPU (x86_64) - Stubbed out
- DirectX 12 - Planned
v0.1.0 Status: Conv2D Gaussian blur available on CUDA + Vulkan backends
Bold = Implemented | Italic = Planned
- Convolution (2D) - Gaussian blur demo
- Bilateral Filter - Edge-preserving smoothing
- Sobel Edge Detection - Computer vision classic
- Histogram Computation - Statistical analysis
- Parallel Reduction - Sum/Max/Min with GPU optimization patterns
- Matrix Transpose - Memory coalescing showcase
- Parallel Prefix Scan - Fundamental GPU algorithm
- N-Body Simulation - Gravitational dynamics
- Monte Carlo Forest Layout - Procedural generation
Getting runtimes working on heterogeneous systems is always the hard part.
Kerntopia's built-in interrogator rivals the best tools:
./kerntopia info --verboseSample Output (Subject to change):
🌾 Kerntopia System Interrogation v0.1.0
CUDA Backend: ✅ Available
├─ Runtime: CUDA 13.0 (/usr/local/cuda/lib64/libcudart.so)
├─ Device 0: NVIDIA GeForce RTX 4060 (SM 8.9, 8GB)
└─ Compiler: nvcc 13.0 (/usr/local/cuda/bin/nvcc)
Vulkan Backend: ✅ Available
├─ Runtime: libvulkan.so.1 (Vulkan 1.3.290)
├─ Device 0: CPU
└─ Compiler: glslangValidator 14.3.0
SLANG Compiler: v2025.14.3 (FetchContent, 234MB)
System: WSL2 Ubuntu 24.04, GCC 13.3.0, x86_64
Want to add your own kernel? Start here:
# Copy the vector_add template
cp -r examples/vector_add my_awesome_kernel
cd my_awesome_kernel
# Edit kernel.slang, update CMakeLists.txt
# Build and run!Future: Complete integration guide in docs/DEVELOPER_GUIDE.md
- STB Libraries - Image I/O (MIT/Public Domain)
- TinyEXR - HDR support (BSD 3-Clause)
- SLANG Compiler - Auto-fetched via CMake
Development: Claude Code and Sonnet Teams
SLANG Platform: Khronos SLANG and Slang Playground developers
Recorded LLM scaffolding (undercount):
- Input/Output Prompts: 48K words, 8.5K lines, 63.8K tokens
- Generated Code: 41.5K words, 12.4K lines, 116.8K tokens
Ready to contribute? Check out our contribution guidelines to get started. We use a Contributor License Agreement (CLA) to ensure the project can evolve while recognizing all contributors.
This project is licensed under the MIT License - see the LICENSE file for details.
While the MIT license only requires copyright notice retention, attribution in benchmark results or commercial usage is appreciated. See ATTRIBUTION.md for guidelines on citing Kerntopia in your work.
Version: 0.1.0
Mission: Democratize GPU compute development through SLANG
Target: GPU developers, SIGGRAPH attendees, GPU enthusiasts
Remember: Every expert was once a beginner. ✨🌾✨