PolyCIM accepted by ICCAD 2026: improving data reuse in digital CIM accelerators with polyhedral compilation

👏 Paper title: PolyCIM: Improving Data Reuse in Digital CIM Accelerators with Polyhedral-Based Compilation.

PolyCIM improves the mapping of modern deep neural network operators onto digital compute-in-memory (CIM) accelerators. Conventional im2col-based mappings mainly exploit axis-aligned data reuse, leaving CIM arrays severely underutilized for operators such as depthwise and grouped convolutions.

The framework uses a unified polyhedral abstraction for DNN workloads and digital CIM architectures. It applies data reuse exposure, computation mapping, and data movement optimization to generate efficient mappings automatically.

PolyCIM is motivated by the observation that data reuse in modern DNN operators often forms hyperplanes along non-axial directions. Affine transformations can realign these hidden reuse patterns with the rows and columns of rigid CIM arrays.

By systematically exposing and mapping these reuse opportunities, PolyCIM provides an efficient compilation path for modern DNN operators on digital CIM accelerators.