Performance Optimization of OpenCL-Based Gaussian Filtering: A Comparative Analysis of Optimization Techniques

Gaussian filtering is a popular image processing algorithm for noise elimination and feature detection. Due to its high computational complexity, real-time processing is difficult, requiring efficient parallel computation. OpenCL, an open-standard parallel programming framework, allows high-performance computing on heterogeneous platforms. This paper provides an optimized OpenCL version of Gaussian filtering by applying performance improvement techniques systematically, such as loop unrolling, vectorization, local memory use, data pack optimization, and producing multiple outputs from each work-item. The effectiveness of the optimizations is tested in terms of execution time using NVIDIA GeForce 920MX with OpenCL 3.0. Experimental outcomes reflect significant improvements in performance and decide the most useful optimization techniques to reduce the latency of execution. These findings are used in optimizing OpenCL kernels to enhance image processing applications. The scope for other optimizations such as auto-tuning and machine learning-based approaches for the future may be explored.