Effective usage of vector registers in decoupled vector architectures

This paper presents a study of the impact of reducing the vector register size in a decoupled vector architecture. In traditional in-order vector architectures, long vector registers have typically been the norm. We start presenting data that shows that, even for highly vectorizable codes, only a small fraction of all elements of a long vector register are actually used. We also show that reducing the register size in a traditional vector architecture in an attempt to reduce hardware cost and maximize register utilization results in a severe performance degradation. However, we combine the decoupling technique with the vector register reduction and show that the resulting architecture tolerates very well the register size cuts. We simulate a selection of Perfect Club and Specfp92 programs using a trace driven approach and compare the execution time in a conventional vector architecture with a decoupled vector architecture using different registers sizes. Halving the register size and using decoupling provides speedups between 1.04-1.49 over a traditional in-order vector machines. Even reducing the register length to 1/4 the original size (and, in some cases, to 1/8) the performance of the decoupled machine is better than a conventional vector model. Moreover, we observe that the resulting decoupled machine with short registers tolerates very well long memory latencies.