SIMD-Aware Ray-Casting

Abstract

With the addition of loop and branch capabilities into the GPU programming set, a more natural and free-flowing volume rendering pipeline execution mode recently emerged. While the execution of the looped fragment program is still SIMD, the numerical approximation of the volume integral is no longer broken up into multiple passes, as it was in previous approaches. Although this certainly is an improvement, due to the less tight control, the present framework lacks the capabilities to focus program flow on the relevant data. In fact, due to the GPU-native SIMD programming model, the slowest ray will now determine the rendering speed. Our paper seeks to provide solutions to enforce more control over the rendering process, with the special SIMD programming model in mind, while retaining this new and promising single-pass rendering paradigm. Our efforts were rewarded with speedups of up to 8.