Technical Papers Archive

Chisa: Dynamic Logical Way Partitioning for Efficient Caching in Sparse Accelerators


Authors: Jiahao Lu (Xiamen University; Institute of Artificial Intelligence, Xiamen University); Zhirong Shen (Xiamen University); and Jiwu Shu (Xiamen University, Tsinghua University)

Abstract: Sparse tensor computations are memory-bound, making on-chip data reuse critical to sparse accelerator performance. Guided cache replacement derives lookahead information from structural relationships between operand tensors, but its effectiveness hinges on the prefetch observation window in SRAM. Under constrained SRAM, the window is inevitably small and guidance becomes noisy or misleading—precisely when caching matters most.

We propose Chisa, a dynamic logical way partitioning scheme for sparse accelerator caches. Inspired by victim caches, Chisa integrates buffer semantics into cache sets: a single role bit lets each way dynamically serve as a logical cache or buffer way without data migration, so mistakenly evicted blocks can be rescued upon reuse. Chisa further incorporates Jaccard similarity-based initialization and rescue rate-based runtime adaptation. On real-world sparse matrices, Chisa achieves geometric mean speedups of 1.23x and 1.27x over the state of the art on SpMSpM and SpMM, respectively, peaking at 4.46x and 4.64x.



Presentation: file

Back to Technical Papers Archive Listing