Energy-Efficient Stream Compaction Through Filtering and Coalescing Accesses in GPGPU Memory Partitions

Graph-based applications are essential in emerging domains such as data analytics or machine learning. Data gathering a knowledge-based society requires great processing efficiency. High-throughput GPGPU architectures key to enable efficient graph processing. Nonetheless, irregular and sparse memory access patterns present graph-based induce high divergence contention, which result poor efficiency for Recent work has pointed out the importance of stream compaction operations, proposed Stream Compaction Unit (SCU) offload them specialized hardware. On other hand, contention caused by been tackled with Irregular accesses Reorder (IRU), delivering improved coalescing. In this paper, we propose new unit, IRU-enhanced SCU (ISCU), that leverages strengths both approaches. The ISCU employs mechanisms IRU improve throughput limitations, achieving synergistic effect We evaluate wide variety state-of-the-art algorithms applications. Results show achieves performance speedup 2.2x 90 percent energy savings derived from reduction 78 accesses, while incurring 8.5 area overhead.