Cross-Filter Structured Pruning for Efficient Sparse CNN Acceleration

초록

Convolutional Neural Networks (CNNs) are widely used in vision tasks for resource-constrained environments due to their computational efficiency and strong generalization. However, the dominance of 1 x 1 convolutions in modern CNN architectures introduces challenges for sparsity-aware hardware accelerators, particularly in processing element (PE) load balancing, which limits speedup in sparse inference. To address this issue, this paper proposes a cross-filter structured pruning method, enforcing a uniform sparsity pattern across multiple filters to ensure balanced workload distribution among PEs. This approach is further extended to k x k convolutions by decomposing them into 1 x 1 filters, improving applicability across various CNN layers. This paper also proposes an intra-kernel parallelism technique which significantly reduces the size of PE's local buffers, a critical bottleneck in sparse CNN accelerators. Experimental results show that the proposed approach maintains accuracy comparable to globally unstructured pruning while significantly enhancing inference speed. FPGA implementation and cycle-accurate simulations confirm improvements in processing speed, energy efficiency, and hardware utilization, making this method well-suited for edge and mobile AI applications. Specifically, the proposed architecture achieves 1.14x to 1.6x speedup over Sparten for various CNN models and delivers 7.6x and 1.9x higher energy efficiency compared to Sparten and StarSPA, respectively. In terms of area efficiency, synthesis results show a 1.73x-10.95x reduction in required hardware primitives compared to Sparten and StarSPA.

키워드