PolyCIM: Improving Data Reuse in Digital CIM Accelerators with Polyhedral-Based Compilation

2026-09-28

AI Summary

Digital compute-in-memory (CIM) accelerators suffer from low array utilization due to structural rigidity that impedes the exploitation of non-axial data reuse. This work proposes a polyhedral-model-based compilation framework that exposes and reorganizes hyperplane data reuse through affine transformations, thereby optimizing computation mapping and data movement. To our knowledge, this study is the first to systematically reveal non-axial hyperplane reuse structures in deep neural networks (DNNs) and establish a unified abstraction adaptable to diverse workloads and architectures. Experimental results demonstrate that the proposed approach improves macro-cell utilization by 4× and achieves a 3.2× speedup, significantly bridging the performance gap between DNN operators and CIM architectures.

Abstract

Digital Compute-in-Memory (CIM) presents a promising solution for accelerating deep neural networks (DNNs) through the integration of computational logic directly within memory arrays. However, mapping modern DNN operators to CIM accelerators often results in severe array underutilization, due to the strict data reuse constraints imposed by the rigid CIM array structure. We observe that data reuse in modern DNNs forms hyperplane structures often oriented along non-axial directions, rendering them invisible to conventional mapping methods that only exploit axis-aligned reuse. In this work, we propose PolyCIM, a polyhedral-based compilation framework for CIM architectures that systematically exposes and realigns these hyperplanes through affine transformations. PolyCIM provides a unified abstraction capable of efficiently representing both diverse DNN workloads and digital CIM architectures. Through data reuse exposure, computation mapping, and data movement optimization, PolyCIM generates mappings for CIM architectures that achieve superior array utilization and performance. Experimental results show that PolyCIM delivers up to $4\times$ improvement in macro utilization and $3.2\times$ speedup, effectively bridging the gap between modern DNN operators and CIM architectures.

Problem

• Compute-in-Memory
• Data Reuse
• Array Underutilization
• Deep Neural Networks
• Operator Mapping

Innovation

• Compute-in-Memory
• Polyhedral Compilation
• Data Reuse
• Affine Transformation
• DNN Accelerator