High-efficiency data loading and output buffering strategy for sparse convolutional computing

In view of the problems such as inefficient data loading, insufficient utilization of multiply-accumulates resources, complex output buffering and addressing logic in existing neural network accelerators when processing sparse neural networks, a high-efficiency data loading and output buffering strategy for sparse convolutional computing was proposed. It performed an all-to-all multiply-accumulates operation on the non-zero input feature map data and the non-zero weights belonging to the same input channel, which reduces the difficulty of non-zero data pairing and improves the utilization of multiply-accumulates resources. By using input stationary calculation and intensive cyclic loading of input feature map data, it significantly reduced the number of data off-chip fetches. It optimized the output buffer design and solved the problems of address access contention and storage congestion during output buffering in existing solutions. Experimental results show that, when compare to fine-grained systolic accelerator with similar architectures, the process element area of the proposed architecture is decreased by 21.45%; the data loading speed is increased by 117.71% on average; the average utilization of multiplier is increased by 11.25%, reaching 89%.