No abstract

No abstract

No abstract

This paper presents a fast and efficient heuristic for pipelining a loop under resource-constraints. The loop is represented as a dependence graph, G , whose nodes are operations that are bound to available resources and edges denote the data dependencies between the operations. The data dependenczes restrict the degree of parallelism that can be achieved whzle scheduling the graph. We propose a fast retimang based graph transformation technaque which relaxes the data dependencies in the graph while maintainzng functional equivalence. Relaxing data dependencaes provzdes more flexibility for the scheduler to schedule operations, thereby leading to faster throughput. Our objective is to obtain a retimed graph which when scheduled achieves an optimal/near-optimal pipelined steady state throughput. A detailed aigorithm is presented t o solve the problem. We provide results that zllustrate the effectaveness of our algorithm.

No abstract