PIMCOMP: An End-to-End DNN Compiler for Processing-In-Memory Accelerators

TL;DR

PIMCOMP is an end-to-end compiler that enables efficient deployment of deep neural networks on processing-in-memory accelerators, optimizing resource utilization and dataflow scheduling for diverse hardware architectures.

Contribution

It introduces a configurable abstraction and multi-level optimization framework for automatic DNN deployment on PIM accelerators, addressing resource and dataflow challenges.

Findings

Improves throughput, latency, and energy efficiency across various PIM architectures.

Supports flexible convolutional layer partitioning and resource mapping.

Enhances system performance through tailored dataflow scheduling algorithms.

Abstract

Various processing-in-memory (PIM) accelerators based on various devices, micro-architectures, and interfaces have been proposed to accelerate deep neural networks (DNNs). How to deploy DNNs onto PIM-based accelerators is the key to explore PIM's high performance and energy efficiency. The scale of DNN models, the diversity of PIM accelerators, and the complexity of deployment are far beyond the human deployment capability. Hence, an automatic deployment methodology is indispensable. In this work, we propose PIMCOMP, an end-to-end DNN compiler tailored for PIM accelerators, achieving efficient deployment of DNN models on PIM hardware. PIMCOMP can adapt to various PIM architectures by using an abstract configurable PIM accelerator template with a set of pseudo-instructions, which is a high-level abstraction of the hardware's fundamental functionalities. Through a generic multi-level optimization framework, PIMCOMP realizes an end-to-end conversion from a high-level DNN description to pseudo-instructions, which can be further converted to specific hardware intrinsics/primitives. The compilation addresses two critical issues in PIM-accelerated inference from a system perspective: resource utilization and dataflow scheduling. PIMCOMP adopts a flexible unfolding format to reshape and partition convolutional layers, adopts a weight-layout guided computation-storage-mapping approach to enhance resource utilization, and balances the system's computation, memory access, and communication characteristics. For dataflow scheduling, we design two scheduling algorithms with different inter-layer pipeline granularities to support varying application scenarios while ensuring high computational parallelism. Experiments demonstrate that PIMCOMP improves throughput, latency, and energy efficiency across various architectures. PIMCOMP is open-sourced at \url{https://github.com/sunxt99/PIMCOMP-NN}.