Enhancing Convolution Computing: An Overlap-Add Fast Fourier Transform Approach

This study underscores the potential of matrix computation transformations to expedite convolution operations in Convolution Neural Networks (CNNs), challenging the runtime efficiencies of traditional convolutional calculations. The prevalent reliance on fixed-point or integer computations has been a notable limitation, affecting the precision of outcomes. By integrating half-precision floating-point standards (IEEE 754) and implementing an Overlap-add Fast Fourier Transform (OA-FFT) methodology, including 8-point DFT computations via the Cooley-Tukey algorithm, this research introduces an enhanced computational framework. Our architecture, designed for performance evaluation and optimization, while resource-intensive, significantly improves runtime and precision. Testing environment and procedure, utilizing the reliable built-in software calculating method as reference for evaluating the efficiency, are also introduced in this work as an easy-to-understand model demonstrating the feasibility of our method. The successful integration of floating-point calculations opens new avenues for accuracy improvements in CNN computational processes, moreover, the potential of hardware accelerating implementation in enhance the effectiveness of algorithms.