Efficient Stride 2 Winograd Convolution Method Using Unified Transformation Matrices on FPGA

Winograd algorithm can effectively reduce the computational complexity of convolution operation. Effectively using the parallelism of Winograd convolution algorithm can effectively improve the performance of accelerator architectures on FPGA. The stride represents the number of elements that the window slides when filter is scanned on the input feature map. The Winograd algorithm with the stride of 2 implemented in previous studies divided the input feature maps into multiple groups of Winograd algorithms to complete the operations, resulting in additional precomputation and hardware resource overhead. In this paper, we propose a new Winograd convolution algorithm with the stride of 2. This method uses the unified Winograd transformation matrices instead of the grouping method to complete the calculation. Therefore, the method proposed in this paper can realize 2D Winograd convolution and 3D Winograd convolution by nested 1D Winograd convolution, just like the Winograd convolution algorithm with the stride of 1. In this paper, Winograd transformation matrices with kernel size of 3, 5, and 7 are provided. In particular, for convolution with the kernel of 3, this method reduces the addition operations of Winograd algorithm by 30.0%-31.5% and removes unnecessary shift operations completely. In addition, we implement Winograd convolution algorithm with the stride of 2 through template design, and realize pipeline and data reuse. Compared to the state-of-the-art implementation, the proposed method results in a speedup of 1.24 and reduces resource usage.