🚀 Spconv Builder & Binary Distribution

Prebuilt, standalone C++ artifacts for spconv — no Python, no pain.

This repository provides a fully automated build & release pipeline that converts traveller59/spconv + FindDefinition/cumm into production-ready C++ headers and shared libraries (libspconv.so).

Designed for high-performance inference environments such as:

• TensorRT custom plugins
• CUDA-native C++ inference engines
• Embedded & autonomous driving platforms (Jetson Orin)

💡 If you’ve ever struggled with “spconv build” — this repo is for you.

---

✨ Why This Exists

spconv is an industry-standard sparse convolution library, but:

• ❌ Strongly coupled with Python build/runtime
• ❌ Painful to integrate into pure C++ inference stacks
• ❌ Hard to cross-compile for Jetson / Drive platforms
• ❌ CI-unfriendly for reproducible builds

This project solves exactly that.

---

🔑 Key Features

✅ Standalone C++ Distribution

• No Python required at runtime or link time
• Clean headers + libspconv.so

✅ Production-Grade Architectures

• x86_64: Data center / workstation GPUs
• aarch64: NVIDIA Jetson AGX Orin / OrinX

✅ Inference-Oriented Build

• Optimized for TensorRT plugin integration
• Suitable for custom CUDA operators

✅ Fully Automated CI

• GitHub Actions builds & publishes binaries
• Versioned, reproducible, downloadable artifacts

✅ Flattened Header Layout

• spconv, cumm, and tensorview headers included
• Ready for direct CMake consumption

---

📦 Prebuilt Artifacts

All binaries are published on the 👉 Releases Page

📁 Package Layout

libspconv-{version}-{arch}-cuda{cuda_version}.tar.gz
├── include/
│   ├── spconvlib/
│   │   ├── cumm/        # CUDA MMA & kernel primitives
│   │   └── spconv/      # Sparse convolution core
│   └── tensorview/      # TensorView utilities
└── lib/
    └── libspconv.so     # Shared library

---

🧩 Typical Use Cases

• 🔌 TensorRT Plugin Development
• 🚗 Autonomous Driving Perception Pipelines
• ⚡ Sparse CNN Inference Engines
• 🧠 CUDA / HPC Research Projects
• 📦 Binary Distribution in CI/CD

---

🛠 Supported Platforms

Architecture	Platform	OS	CUDA	Compute Capability
x86_64	PC / Server	Ubuntu 20.04	11.8	8.0 / 8.6 / 8.9
aarch64	Jetson Orin / OrinX	Ubuntu 20.04 (JetPack 5.x / DriveOS 6.0+)	11.4	8.7

⚠️ Other CUDA / OS versions may work but are not officially validated.

---

🧠 Design Philosophy

• Inference-first, not training
• C++ native, not Python-bound
• Reproducible builds, not local hacks
• CI-driven, not manual compilation

---

📜 License & Acknowledgements

This repository does not modify the original algorithms.

• Build & CI Scripts: MIT License
• spconv: Apache 2.0 — traveller59
• cumm: Apache 2.0 — FindDefinition

🙏 Full credit to the original authors.

If you use this in research or production, please cite the original spconv project.

---

⭐ Star This Repo If It Helped You

If this saved you days of build pain or unblocked your inference pipeline, a star is the best way to say thanks ⭐

---

🚧 Roadmap

- [ ] TensorRT plugin examples
- [ ] CMake config package (find_package support)
- [ ] CUDA 12.x support