Parallel Hierarchical Radiosity On Cache-Coherent Multiprocessors

Computing radiosity is a computationally very expensive problem in computer graphics. Recent hierarchical methods have greatly speeded up the computation of first diffuse and now also specular radiosity. We present a parallel algorithm for computing both diffuse and specular radiosity together, and examine its performance in detail on cache-coherent shared address space multiprocessors. We compare this with an earlier implementation of parallel diffuse-only radiosity computation, which has very different characteristics. The algorithm is both irregular and highly unpredictable. Despite this, by taking advantage of the cache-coherent shared address space and using distributed dynamic task queues for load balancing, we obtain speedups of 26.7 on a 32-processor machine with distributed memory, with no need for explicit data management, and 14.2 on a 16-processor machine with centralized memory. Both the diffuse and the specular+diffuse programs achieve very good load balance. Because the specular program must process many more source patches for each destination patch than does the diffuse program, it has greater memory stall time; however the memory access overheads do not increase with the number of processes or lower the speedups attained. Detailed performance debugging was used to identify and understand bottlenecks in the programs, which mostly have to do with contention for locks that protect shared data structures, albeit in a very fine-grained way. We conclude that moderate-scale shared address space multiprocessors provide a successful platform for speeding up hierarchical radiosity computations, both diffuse and specular.