Stream data prefetcher for the GPU memory interface

Data caches are often unable to efficiently cope with the massive and simultaneous requests imposed by the SIMT execution model of modern GPUs. While software-aided cache management techniques and scheduling approaches were early considered, efficient prefetching schemes are regarded as the most viable solution to improve the efficiency of the GPU memory subsystem. Accordingly, a new GPU prefetching mechanism is proposed, by extending the stream computing model beyond the actual GPU processing core, thus broadening it toward the memory interface. The proposed prefetcher takes advantage of the available cache management resources and combines a low-profile architecture with a dedicated pattern descriptor specification, which is used to explicitly encode each kernel memory access pattern. The obtained results show that the proposed mechanism increases the L1 data cache hit rate by an average of 61%, resulting in performance speedups as high as 9.2 and consequent energy efficiency improvements as high as 11.