Programming Approach for Accelerator-based Heterogeneous Multiprocessors for Scientific Applications

We address programming of accelerator-based heterogeneous multiprocessors in the context of computational science. Specifically, we consider stream architectures with explicitly man-aged memory hierarchies. In this paper we present a programming approach which supports program development for such multiprocessors. The programming approach is based on a coordination model which allows a programmer explicitly to control parallel activities and to manage memory hierarchies. Accelerators are only beneficial, if one succeeds to map the computational kernel efficiently onto the non-general-purpose hardware. Since the tar-get architecture of our programming system are stream multiprocessors, namely Cell/BE, streaming abstractions are provided to improve programmability of stream kernels and to enable profitable compiler optimizations. Parallelization techniques for code generation are presented. First experiences back up the approach.