Loop Scheduling Optimization with Data Prefetching based on Multi-dimensional Retiming

In this paper, we propose a novel loop scheduling technique based on multi-dimensional retiming in a balanced fashion, which considers the computation schedule and memory access schedule simultaneously. Experiments show that the proposed technique which combines the data prefetching and retiming is successful in hiding memory latency and improving the overall performance. Comparing with the traditional list scheduling algorithm, the average improvement is 28%. 1 Introduction In the past few years, the processing time of a processor has increased more rapidly than the speed to access the memory. In a system with a local and remote memory, the program execution time become significantly depending on the remote-memory access latency. Data prefetching, i.e., retrieving data from the remote memory and store it in the local memory before using it, is an attractive approach to hide memory access latencies. A prefetching scheme reduces the memory miss penalty by overlapping the processor comp..