Dual-precision fixed-point arithmetic for low-power ray-triangle intersections

摘要

Ray-Triangle intersection is a fundamental computation in most ray tracing algorithms. The prohibitive cost of the ray-triangle test algorithms limits the utilization of these algorithms in settings with low power budgets, such as mobile systems. In this work, we analyze the precision requirements for ray-triangle intersection and observe that for most of the rays a low fixed-point precision is sufficient and only for a small fraction of rays a high precision is required. Accordingly, we propose a dual-precision fixed-point hardware accelerator for ray-triangle intersection targeting low-power systems, where the higher precision is only activated for tests deemed critical by our algorithm. Towards this goal, we develop a thresholding technique that autonomously switches between the lower and higher precisions, where the lower precision unit is used for the majority of the tests resulting in significant benefits in power consumption. We evaluate our methodology on a representative set of scenes and our proposed methodology in hardware. Our methodology introduces negligible accuracy errors in ray-triangle tests with an average of 0.25% misclassifications, while offering 87% savings in energy consumption compared to a baseline floating-point design and 27% savings compared to a high-precision fixed-point design. (C) 2020 Elsevier Ltd. All rights reserved.