An Im2col Architecture Using The Benes Network For Deep Learning Hardware Accelerators

General matrix-matrix multiply (GEMM) has a significant effect on the performance of deep learning hardware accelerators. Image to column (im2col) is a process that rearranges input data images to facilitate efficient GEMM operation. This paper proposes a method for implementing the im2col process using a benes network. The proposed method eliminates bottlenecks occurring in the im2col process and achieves a 33% reduction in on-chip memory usage. Additionally, depending on the routing algorithm of the benes network, images can be reorganized into the desired configuration.