Evaluating Winograd Algorithm for Convolution Neural Network using Verilog

Convolutional neural networks (CNNs) are very resource intensive and consume a lot of computational power. The convolution operation itself is a very complex process. Hence this work deals with optimizing the convolution operation by employing an efficient algorithm called the “Winograd algorithm”. The performance of the algorithm was evaluated and compared to that of the general matrix multiplication method (GeMM) of the convolution. The RTL models of GeMM and Winograd algorithm were synthesized on the Xilinx XCVU9P FPGA. The synthesis results show that Winograd algorithm was resource efficient and utilizes 42.98% lesser resources than the GeMM algorithm. The power consumption for Winograd algorithm is 39.47% lesser than that of GeMM.