Cache-aware sampling strategies for texture-based ray casting on GPU

As a major component of volume rendering, the ray casting algorithm is memory-intensive. However, most existing texture-based volume rendering methods blindly map computational resources to texture memory and result in an incoherent access pattern, causing low cache hit rates in certain cases. The distance between samples taken by threads of the same scheduling unit (e.g. a warp of 32 threads in CUDA), of the GPU is a major factor that affects the texture cache hit rate. Based on this fact, we present a new sampling strategy, i.e. warp marching, which displays a novel computation-to-core mapping. In addition, a double buffer approach is introduced and special GPU operations are leveraged to improve the efficiency of parallel executions. To keep a roughly constant rendering performance when rotating the volume, we change our warp marching algorithm, so that samples can be taken along different directions of the volume. As a result, varying texture cache hit rates in different viewing directions are averaged out. Through a series of micro-benchmarking and real-life data experiments, we rigorously analyze our sampling strategies, and demonstrate significant performance enhancements over existing sampling methods.