LooPIM: A Loop-Oriented Acceleration Framework for Processing-in-Memory

Abstract

The latest development of 3D stacking technology provides a feasible solution for processing in memory (PIM) to solve the memory wall crisis. Since data can be processed locally in a logic layer adjacent to memory, large emerging big data applications may achieve significant performance and energy-efficiency benefits. Researchers continue to propose various PIM architectures to utilize the advantages of stacked memory efficiently. However, previous efforts relied on manually mapping specialized kernels to the logic layer, making it infeasible to perform more general workloads. We propose a loop-oriented acceleration framework that intelligently partitions loops and maps them onto appropriate execution units. The mapping mechanism can significantly reduce the communication overhead between the host processor and the PIM stack. Besides, we introduce a simple execution model that exploits data-level parallelism transparently. Experiments show that our framework achieves approximately 2.4x performance speedups comparing to the baseline 3D memory system.