Comparison of Highly Configurable CPU- and GPU-based Convolution Engines

In this work the performance of real-time audio signal processing convolution engines is evaluated. A CPU-based implementation using the Integrated Performance Primitives Library and two GPU-based implementations using CUDA and OpenCL are compared. The purpose of these convolution engines is auralization, e.g., the binaural rendering of virtual multichannel configurations. Any multichannel input and output configuration is supported, e.g., 22.2 to 5.1, 7.1 to 2.0, vice versa, etc. This ability results in a trade-off between configurability and performance. Using a 5.1-to-binaural setup with continuous filter changes due to simulated head-tracking, GPU processing is more efficient when 24 filters of more than 1.92 seconds duration each @ 48 kHz sampling rate are convolved. The GPU is capable of convolving longer filters in real-time than a CPU-based processing. By comparing both GPU-based implementations, negligible performance differences between OpenCL and CUDA were measured.