GPU Connected Components Benchmark

Benchmarking connected component labeling libraries across CPU and GPU implementations.

Suggested to me by my advisor @tvercaut

Benchmarked Libraries (Native 3D Only)

Library	Backend	3D Support
PyTorch (cuCIM)	GPU	Native 3D
CuPy ndimage	GPU	Native 3D
cuCIM	GPU	Native 3D
cc3d	CPU	Native 3D
scikit-image	CPU	Native 3D
SimpleITK	CPU	Native 3D
Mahotas	CPU	Native 3D
DIPlib	CPU	Native 3D

Results

Installation

Requires Python 3.9-3.12.

# Using uv (recommended)
uv venv --python 3.11
uv sync

# Install GPU support (CUDA 11.x)
uv pip install torch --index-url https://download.pytorch.org/whl/cu118
uv pip install cupy-cuda11x cucim

# OR for CUDA 12.x
uv pip install torch --index-url https://download.pytorch.org/whl/cu121
uv pip install cupy-cuda12x cucim

Usage

# Run benchmark via CLI
uv run benchmark-cc

# Or as module
uv run python -m gpu_cc.benchmark

Project Structure

├── src/
│   └── gpu_cc/
│       ├── __init__.py
│       ├── gpu_wrapper.py
│       └── benchmark.py
├── plots/
├── pyproject.toml
├── uv.lock
└── README.md

Dependencies

Base (CPU, native 3D only):

• numpy, matplotlib
• scikit-image, connected-components-3d (cc3d)
• SimpleITK, mahotas, diplib

GPU (native 3D only):

• torch (with CUDA)
• cupy-cuda11x or cupy-cuda12x
• cucim