CUDA Kernel Academy

From Zero to Hero: Systematic CUDA High-Performance Kernel Development
从零到极致：系统性学习 CUDA 高性能算子开发

English | 简体中文

---

A structured CUDA learning repository covering matrix multiplication, reusable kernels, advanced optimization techniques, and a lightweight inference engine. Master GPU programming from SGEMM basics to Tensor Core optimization.

Features

Feature	Description
Progressive Learning Path	4 interconnected modules from basics to production
Performance-Focused	Real benchmarks against cuBLAS, not toy examples
Modern C++	Leverages C++17/20 features for clean, safe GPU code
Production Patterns	Header-only library design, memory pools, stream management
Multi-Architecture	Supports Volta (sm_70) through Hopper (sm_90)

Documentation

• Installation Guide
• Coding Style
• Integration Examples
• FAQ
• Troubleshooting
• Contributing

Sub-projects

#	Project	Focus	Build
01	SGEMM Tutorial	Progressive SGEMM optimization	Standalone Makefile
02	TensorCraft Core	Header-only kernel library	CMake
03	HPC Advanced	Advanced CUDA/HPC techniques	CMake
04	Inference Engine	Lightweight DL inference engine	CMake

Learning path

01-SGEMM Tutorial (1-2 weeks)
        ↓  Master shared memory, bank conflicts, WMMA
02-TensorCraft Core (2-3 weeks)
        ↓  Build reusable kernels, API design
03-HPC Advanced (3-4 weeks)
        ↓  CUDA 13 features, FlashAttention
04-Inference Engine (2-3 weeks)
        ↓  Complete inference framework

Prerequisites: C/C++ basics, linear algebra fundamentals. CUDA experience helpful but not required.

Quick start

git clone https://github.com/LessUp/cuda-kernel-academy.git
cd cuda-kernel-academy

cmake --preset default
cmake --build --preset default
ctest --preset default

List available presets with:

cmake --list-presets

Build notes

• The root CMake build covers 02-tensorcraft-core, 03-hpc-advanced, 04-inference-engine, common, and examples.
• 01-sgemm-tutorial is intentionally standalone and uses its own Makefile.
• GitHub Actions currently runs CPU-safe checks (formatting, docs, links, preset validation). Full CUDA builds/tests should be run on a local machine with a GPU.

Build options

Option	Default	Description
BUILD_TENSORCRAFT	ON	Build TensorCraft Core
BUILD_HPC_ADVANCED	ON	Build HPC Advanced
BUILD_INFERENCE_ENGINE	ON	Build Inference Engine
BUILD_EXAMPLES	ON	Build examples
BUILD_TESTS	ON	Build tests
BUILD_BENCHMARKS	ON	Build benchmarks
BUILD_PYTHON_BINDINGS	OFF	Build optional Python bindings

Requirements

Component	Minimum	Recommended
CUDA Toolkit	11.0	12.x
CMake	3.20	3.24+
Compiler	GCC 9 / Clang 10	GCC 11+
GPU	Volta (sm_70)	Ampere/Ada (sm_80+)

Supported Architectures:

Arch	sm	GPUs
Volta	70	V100
Turing	75	RTX 2080, T4
Ampere	80, 86	A100, RTX 3090
Ada	89	RTX 4090, L40
Hopper	90	H100

References

• CUDA C++ Programming Guide
• CUTLASS - CUDA Templates for Linear Algebra
• Simon Boehm's GEMM Tutorial - Excellent optimization walkthrough
• NVIDIA Developer Blog - Latest techniques and best practices

Citation

If you find this project helpful in your research or work:

@misc{cuda-kernel-academy,
  author = {CUDA Kernel Academy Contributors},
  title = {CUDA Kernel Academy: A Comprehensive Learning Path for High-Performance CUDA Kernel Development},
  year = {2026},
  publisher = {GitHub},
  url = {https://github.com/LessUp/cuda-kernel-academy}
}

License

MIT License