Abstract: Over the past few years, the powerful computation rates and high memory bandwidth of GPUs have attracted efforts to run raytracing on GPUs. Our work extends Foley et al.'s GPU k-d tree research. We port their kd-restart algorithm from multi-pass, using CPU load balancing, to single pass, using current GPUs' branching and looping abilities. We introduce three optimizations: a packetized formulation, a technique for restarting partially down the tree instead of at the root, and a small, fixed-size stack that is checked before resorting to restart. Our optimized implementation achieves 15 - 18 million primary rays per second and 16 - 27 million shadow rays per second on our test scenes.

Our system also takes advantage of GPUs' strengths at rasterization and shading to offer a mode where rasterization replaces eye ray scene intersection, and primary hits and local shading are produced with standard Direct3D code. For 1024×1024 renderings of our scenes with shadows and Phong shading, we achieve 12-18 frames per second. Finally, we investigate the efficiency of our implementation relative to the computational resources of our GPUs and also compare it against conventional CPUs and the Cell processor, which both have been shown to raytrace well.