high performance computing on graphics processing units: hgpu.org

Rendering methods based on ray tracing provide high image realism, but have been historically regarded as offline only. This has changed in the past decade, due to significant advances in the construction and traversal performance of acceleration structures and the efficient use of data-parallel processing. Today, all major graphics companies offer real-time ray tracing solutions. The following work has contributed to this development with some key insights. We first address the limited support of dynamic scenes in previous work, by proposing two new parallel-friendly construction algorithms for KD-trees and BVHs. By approximating the cost function, we accelerate construction by up to an order of magnitude (especially for BVHs), at the expense of only tiny degradation to traversal performance. For the static portions of the scene, we also address the topic of creating the "perfect" acceleration structure. We develop a polynomial time non-greedy BVH construction algorithm. We then modify it to produce a new type of acceleration structure that inherits both the high performance of KD-trees and the small size of BVHs. Finally, we focus on bringing real-time ray tracing to commodity desktop computers. We develop several new KD-tree and BVH traversal algorithms specifically tailored for the GPU. With them, we show for the first time that GPU ray tracing is indeed feasible, and it can outperform CPU ray tracing by almost an order of magnitude, even on large CAD models.