CoopRT: Accelerating BVH Traversal for Ray Tracing via Cooperative Threads

Ray Tracing is a rendering technique to simulate the way light interacts with objects to create realistic images. It has become prominent thanks to the latest hardware support on Graphics Processing Units (GPUs), i.e., the Ray-Tracing (RT) unit, specially designed to accelerate ray tracing operations. Despite such hardware advances, ray tracing remains a performance bottleneck for high-performance graphics workloads, such as real-time path tracing (PT), which is an application of ray tracing where multiple bouncing rays are traced per pixel. The key reasons are (a) the costly Bounding Volume Hierarchy (BVH) traversal operation, and (b) low Single-Instruction-Multiple-Thread (SIMT) efficiency as the rays in the same warp deviate inevitably in their traversal paths.

In this work, we propose a novel architecture design for cooperative BVH traversal that exploits the parallelism present in the BVH traversal process. The key idea of our CoopRT scheme is to make use of the idle threads, either completely inactive when the ray tracing instruction is executed or partially idle due to early completion, to help the long running threads in the same warp. Specifically, we enable idle threads in a GPU warp to utilize their readily available traversal hardware to help traverse the BVH tree for the busy threads, therefore helping them finish their traversal much faster. This approach is implemented purely in hardware, requiring no changes to the programming model. We present our architecture design and show that it only involves small changes to the existing RT unit.

We evaluated CoopRT in Vulkan-sim, a cycle-level simulator, and observed up to 5.11x speedup over the baseline, with a geometric mean of 2.15x speedup at the cost of a moderate area overhead of 3.0% of the warp buffer in the RT unit. Using the energy-delay product, our CoopRT achieves an average of 2.29x improvement over the baseline.