Depthwise Convolution with Winograd Algorithm for Efficient CNN Acceleration

This work explores an efficient convolutional acceleration framework tailored for edge devices by integrating Depthwise Convolution with the Winograd algorithm. Through RTL-based hardware implementation, the proposed architecture achieves a 2.17× speedup over conventional CNNs, demonstrating its potential for real-time, energy-efficient inference. On the software side, experiments on the CIFAR-100 dataset evaluate Depthwise Convolution with varying kernel sizes (3×3, 5×5, 7×7), revealing that larger kernels may degrade classification accuracy despite reducing operation counts. Additionally, while Winograd F(2,3) maintains accuracy and latency benefits in depthwise settings, Winograd F(4,3) introduces numerical instability and performance loss. These findings underscore the importance of aligning algorithmic optimizations with architectural constraints, particularly for resource-constrained environments and small-resolution inputs..