Computer systems are increasingly parallel and heterogeneous, while programs are still largely written in sequential languages. The obvious suggestion that the compiler should automatically distribute a sequential program across the system usually fails in practice because of the complexity of dependence analysis in the presence of aliasing. We introduce the sieve language construct which facilitates dependence analysis by using the programmer’s knowledge about data dependences and makes code more amenable to automatic parallelisation. The behaviour of sieve programs is deterministic, hence predictable and repeatable. Commercial implementations by Codeplay shows that sieve programs can be efficiently mapped onto a range of systems. This suggests that the sieve construct can be used for building reliable, portable and efficient software for multi-core systems.
展开▼
