RTSpMSpM: Harnessing Ray Tracing for Efficient Sparse Matrix Computations

This repository contains the code and benchmark suite for RTSpMSpM, a novel approach that leverages NVIDIA’s hardware-accelerated ray tracing (RT Cores) to speed up Sparse Matrix × Sparse Matrix Multiplication (SpMSpM). This project demonstrates the feasibility and benefits of mapping sparse matrix operations to the ray tracing pipeline.

Technologies Used

• Languages: C++, Python
• GPU Frameworks: CUDA 12.3, NVIDIA OptiX 8.0.0, cuSPARSE
• Build Tools: CMake 3.22, GCC 7.5.0
• Containers: Docker 27.3.1 with NVIDIA support
• Datasets: SuiteSparse Matrix Collection

Project Structure

RTSpMSpM/
├── cuSparse/                # GPU baseline using cuSPARSE
├── Dockerfile/              # Docker build scripts
├── optixSpMSpM/             # OptiX SDK and build system
│   ├── build/               # Compiled binaries and CMake output
│   └── src/            
│       ├── data/            # Input Datasets
│       ├── support/
│       ├── sutil/     
│       └── optixSpMSpM/     # Core ray tracing-based SpMSpM logic
└── Tool/
    └── Script/
        ├── AE_test.py               # Main script to launch experiments and benchmark
        ├── install.sh               # Compile program
        └── download_dataset.sh      # Benchmark automation script

Installation and Setup

Step 1: Clone the Repository

git clone https://github.com/escalab/RTSpMSpM.git
cd RTSpMSpM

Step 2: Build the Docker Image

cd Dockerfile
./build_image.sh

Step 3: Start the Docker Container

./start_image.sh

Step 4: Enter the Docker Container

./run.sh

Step 5: Install and Compile Inside the Container

cd RTSpMSpM/scripts
./install.sh

How to Run

If runned the installed script, skip to step 3: To run the experiment

To compile the GPU Baseline:

cd /home/RTSpMSpM/cuSparse/src
make

To compile the RT-based SpMSpM implementation:

mkdir -p /home/RTSpMSpM/optixSpMSpM/build
cd /home/RTSpMSpM/optixSpMSpM/build
cmake ../src
make

To run the experiment:

python3 /home/RTSpMSpM/scripts/AE_test.py

6. Artifact Details

• Artifact Availability: Public
  📦 Zenodo Archive
  💻 GitHub Repo

• Expected Output:
  Execution time (latency in milliseconds) for sparse matrix benchmarks, shown in logs or console output.

• Evaluation Time:

  • Setup: ~10 minutes
  • Experiment Runtime: 2–3 hours

• Hardware Requirements:

  • GPU: NVIDIA GPU with compute capability 5.0+ (7.5 recommended)
  • CPU: Original Hardware Intel Core i7 14700K
  • RAM: Original Hardware 128GB DDR4
  • Disk Space: up to ~2GB per datasets

• Software Requirements:

  • CUDA 12.3
  • Docker 27.3.1 (or cmake 3.22 + gcc 7.5.0 if building natively)
  • nvidia-docker recommended for easy setup

• Licensing:

  • Code: MIT License
  • Datasets: Original SuiteSparse licenses

---

7. Citation

If you use this artifact in your research, please cite the corresponding ISCA 2025 paper:

RT+SpMSpM: Harnessing Ray Tracing for Efficient Sparse Matrix Computations, ISCA 2025.