Analysis of CPU and GPU implementations of convolution reverb effect

Abstract

Rapid development of modern central processing units (CPUs) and graphics processing units (GPUs) has allowed a significant increase in computing power for different engineering applications. Audio signal processing is an example of such a computationally demanding application. Fast Fourier Transform (FFT) is often a core part of these processing algorithms, and it is efficiently implemented on the CPUs and GPUs through available libraries. In this paper, we present an implementation of the convolution reverb effect using OpenMP and FFTW library on the CPU, and CUDA and cuFFT library on the GPU. Implemented effect is tested with the set of four different audio signals and ten impulse responses of different lengths. We observed speedups in the range of 2 to 3 times over CPU implementation. The results of the analysis are briefly discussed with emphasis on the benefits and drawbacks of using GPUs in such an application.