Doubly-Block Circulant Kernel Matrix Exploitation in Convolutional Accelerators.

In this paper, we present a novel algorithmic and hardware co-design approach specifically tailored for efficient 2D convolution implementations, a crucial operation in convolutional neural networks (CNNs). Our method addresses the limitations of existing software-based solutions and hardware-based architectures, delivering significant improvements in asymptotic behavior for generic convolution cases. By leveraging the distinctive geometry of doubly block circulant unrolled kernel matrices, our approach eliminates the need for input and weight buffers, optimizes output memory usage, and minimizes redundant memory accesses. A comprehensive comparative analysis with state-of-the-art techniques showcases the key advantages and superior performance of our proposed method, achieving substantial reductions in memory requirements and high throughput.