Graphics processing units (GPUs) provide both memory bandwidth and arithmetic performance far greater than that available on CPUs but, because of their Single-Instruction-Multiple-Data (SIMD) architecture, they are hard to program. Most of the programs ported to GPUs thus far use traditional data-level parallelism, performing only operations that operate uniformly over vectors.NESL is a first-order functional language that was designed to allow programmers to write irregular-parallel programs -such as parallel divide-and-conquer algorithms -for wide-vector parallel computers. This paper presents our port of the NESL implementation to work on GPUs and provides empirical evidence that nested data-parallelism (NDP) on GPUs significantly outperforms CPUbased implementations and matches or beats newer GPU languages that support only flat parallelism. While our performance does not match that of hand-tuned CUDA programs, we argue that the notational conciseness of NESL is worth the loss in performance. This work provides the first language implementation that directly supports NDP on a GPU.

Data parallelism has proven to be an effective technique for highlevel programming of a certain class of parallel applications, but it is not well suited to irregular parallel computations. Blelloch and others proposed nested data parallelism (NDP) as a language mechanism for programming irregular parallel applications in a declarative data-parallel style. The key to this approach is a compiler transformation that flattens the NDP computation and data structures into a form that can be executed efficiently on a widevector SIMD architecture. Unfortunately, this technique is ill suited to execution on today's multicore machines. We present a new technique, called data-only flattening, for the compilation of NDP, which is suitable for multicore architectures. Data-only flattening transforms nested data structures in order to expose programs to various optimizations while leaving control structures intact. We present a formal semantics of data-only flattening in a core language with a rewriting system. We demonstrate the effectiveness of this technique in the Parallel ML implementation and we report encouraging experimental results across various benchmark applications.