Back to Home

Multi-Threaded Vision Pipeline

This project implements a Thread Pool architecture to perform parallel Sobel edge detection. It is designed to maximize CPU throughput while minimizing lock contention and avoiding false sharing at thread boundaries.

Performance: 3.83x speedup
C++17Thread PoolSobelCMake
GitHub

Key Engineering Features

  • Zero-Copy Memory Management: Image buffers are managed via std::shared_ptr<std::vector<uint8_t>>. Data is passed to worker threads by reference, eliminating expensive memory allocations during frame handoffs.
  • Spatial Partitioning & Halo Regions: Workloads are horizontally partitioned based on std::thread::hardware_concurrency(). The architecture implements +1/-1 pixel “Halo Regions” (ghost zones) to allow threads to read overlapping boundary pixels without crossing strict memory bounds, preserving L1/L2 Cache locality and preventing cache thrashing.
  • Custom Thread Pool: Utilizes a fixed-size pool of worker threads synchronized via std::condition_variable and std::mutex, completely avoiding the overhead of thread creation/destruction per frame.