Optimization of the Modified Gaussian Filter for Mobile GPU Usage in Game Workloads

The harnessing of mobile Graphics Processing Units (mGPUs) has become increasingly significant in modern computing technology, especially for mobile devices such as smartphones and tablets. In recent years, however, the focus of research has shifted from graphics-specific studies to those involving General-Purpose computing on Graphics Processing Units (GPGPU), with an emphasis on AI workloads. As a result, there has been a lack of progress in GPU research tailored to the needs of mobile GPU-intensive workloads, such as 3D, VR, and AR gaming. To address this issue, we analyze gaming workloads, which demand significant GPU usage compared to other applications, to identify the specific constraints they impose on mobile GPUs. Subsequently, we investigate optimization strategies that can be applied to address these constraints while ensuring efficient performance for mobile gaming workloads. We provide a detailed breakdown of the differences in shader usage and filter values between game and general workloads. Our observations show that game workloads mostly make use of Linear Filter instead of Nearest Filter, which is widely used in common mobile apps. Although Linear Filters are more accurate than other filters, the Peak Signal-to-Noise Ratio (PSNR) evaluation metric reveals that they are not the most optimal solution for efficiency in game apps. Based on this observation, we present a new filter, called Gaussian ⁺ , optimized for game workloads in mobile GPU systems. We show that Gaussian+ achieves up to 38% improvement in performance with minimal accuracy loss compared to the any other alternate filters.