CUDA-PIM: End-to-End Integration of Digital Processing-in-Memory from High-Level C++ to Microarchitectural Design

Digital processing-in-memory (PIM) architectures mitigate the memory wall problem by facilitating parallel bitwise operations directly within memory. Recent works have demonstrated their algorithmic potential for accelerating data-intensive applications; however, there remains a significant gap in the programming model and microarchitectural design. This is further exacerbated by the emerging model of partitions, which significantly complicates control and periphery. Therefore, inspired by NVIDIA CUDA, this paper provides an end-to-end architectural integration of digital memristive PIM from an abstract high-level C++ programming interface for vector operations to the low-level microarchitecture. We begin by proposing an efficient microarchitecture and instruction set architecture (ISA) that bridge the gap between the low-level control periphery and an abstraction of PIM parallelism into warps and threads. We subsequently propose a PIM compilation library that converts high-level C++ to ISA instructions, and a PIM driver that translates ISA instructions into PIM micro-operations. This drastically simplifies the development of PIM applications and enables PIM integration within larger existing C++ CPU/GPU programs for heterogeneous computing with significant ease. Lastly, we present an efficient GPU-accelerated simulator for the proposed PIM microarchitecture. Although slower than a theoretical PIM chip, this simulator provides an accessible platform for developers to start executing and debugging PIM algorithms. To validate our approach, we implement state-of-the-art matrix operations and FFT PIM-based algorithms as case studies. These examples demonstrate drastically simplified development without compromising performance, showing the potential and significance of CUDA-PIM.