Parallel image pre-processing for in-game object classification

Games that involve photo-realistic rendering of virtual worlds, Special effects, VR, involve highly complex calculations on the GPU (Graphics Processing Unit). While GPUs have specialized cores that accelerate Graphics operations, they are also capable of general purpose computing. In this paper, a specific data flow is chosen for the use-case of in-game object-classification. This use-case involves converting large input graphics resolutions (4k) to much lower resolutions (256×256) needed for compute. We compare the different approaches available for executing the data flow using CUDA (Compute Unified Device Architecture) without impacting gaming performance, and publish comparisons on different classes of GPUs for the first time. It is shown that best performance is achieved with a combination of well optimized algorithms, priority of work assignment, and a flexible execution framework.