high performance computing on graphics processing units: hgpu.org

Bidirectional Path Tracing (BDPT) is a robust unbiased rendering algorithm that samples paths by connecting eye and light paths. By optimally combining different sampling strategies using Multiple Importance Sampling (MIS), BDPT efficiently renders scenes with complex light effects. However, BDPT does not map well on a streaming architecture such as the GPU; Stochastic path lengths lead to an uneven workload between BDPT samples and a large memory footprint that depends on the maximum allowed path length. In this paper, we present a streaming alternative to BDPT, called SBDPT. We present a novel scheme for recursively computing balance heuristic MIS weights. This scheme allows for bidirectional connections to be evaluated independently, only requiring data from the light and eye vertex directly involved in the connection. Using this property, we modify the BDPT algorithm to generate a new eye path for each light vertex. The resulting SBDPT algorithm only requires temporary storage for the last vertex on both eye and light paths. Our implementation achieves high performance and has a fixed low memory footprint, independent of the maximum path length.