Maximizing GFLOPS on Mobile GPUs: Analyzing OpenCL Kernel Optimizations on Snapdragon and MediaTek Platforms

The increasing demand for mobile video and high-resolution image processing demands efficient utilization of GPU computing capability. Mobile GPUs such as ARM Mali and Qualcomm Adreno are crucial in accelerating image and video processing tasks where floating-point computational throughput plays an important role in overall performance. This work provides an empirical study of optimization methods for maximizing GFLOPS on mobile GPUs, with a focus on Snapdragon SM8650 (Adreno 750) and MediaTek MTK6886 (Mali-G610). Different optimization methods, such as vectorization, loop unrolling, and floating-point operations per work item maximization, are implemented and compared. The experiments show significant differences in optimization performance across architectures, reflecting the architectural limitations and performance bottlenecks of each GPU. The results shed light on how to optimize GPU compute workloads for real-time image and video processing applications, which allows for improved efficiency in computational photography, video post-processing, and AI-based multimedia processing.