| dc.description.abstract | This paper studies the problem of instruction assignment and scheduling on spatial architectures. Spatial architectures are architectures whose resources are organized in clusters, with non-zero communication delays between the clusters. On these architectures, instruction scheduling include both space scheduling, where instructions are mapped to clusters, and the traditional time scheduling. This paper considers the problem from both the theoretical and practical perspectives. It presents two integer linear program formulations with known performance bounds. We also present an 8-approximation algorithm for constant m and constant communication delays. Then, we introduce three heuristic algorithms based on list scheduling. Then we study a layer partitioning method. Our final algorithm is a combination of layer partitioning and the third heuristic. Two of the better algorithms are evaluated on the Raw machine. Results show that they are competitive with previously published results; for scientfici codes, our heuristics can perform an average of 25% better. | en_US |
