Memory Efficient Parallel Ray-casting Algorithm for Unstructured Grid Volume Rendering

Abstract

We propose a novel memory efficient parallel ray casting algorithm for unstructured grid. To reduce the high memory consumption of a previous work (i.e. Bunyk), we use a small size of local buffer for each thread to keep view dependent information for most recently visited faces. To improve the utilization efficiency of a local buffer, we propose an index-based hash function and a novel group traversal scheme. With our method, a small size buffer achieves a high hit ratio (i.e. utilization efficiency). As a result, we achieved a compatible performance with Bunyk while using less than 1% of memory space for view dependent face information. Also, our method shows even better performance than Bunyk for a large dataset since the buffer size is small enough to utilize CPU caches and our traversal scheme maximally takes advantage of the early ray termination optimization.