Heterogeneous real-time multi-channel time-domain feature extraction using parallel sum reduction on GPU

Arnin, J. and Kahani, D. and Lakany, H. and Conway, B. A.; (2019) Heterogeneous real-time multi-channel time-domain feature extraction using parallel sum reduction on GPU. In: Proceedings of the 8th Graz Brain Computer Interface Conference 2019, Bridging Science and Application. TUGraz Digital Library, USA. ISBN 9783851256826

[img]
Preview
Text (Arnin-etal-BCIC-2019-Heterogeneous-real-time-multi-channel-time-domain-feature-extraction-using-parallel-sum-reduction-on-GPU)
Arnin_etal_BCIC_2019_Heterogeneous_real_time_multi_channel_time_domain_feature_extraction_using_parallel_sum_reduction_on_GPU.pdf
Final Published Version
License: Creative Commons Attribution-NonCommercial 4.0 logo

Download (584kB)| Preview

    Abstract

    Online BCI has become a fascinating field of research nowadays. One of the main challenges in this field is to reduce the latency caused by the computational complexity of the signal processing algorithms. This issue leads to difficulty in processing real-time data. Usually, a trade-off needs to be considered between the number of input samples and precision of the processing algorithms. In this paper, heterogeneous computing concept is investigated to alleviate the computational complexity occurred in real-time processing. An OpenCL was utilized to implement signal processing algorithms in parallel. Feature extraction methods including band power and statistical moments were selected to examine the power of heterogeneous computing using parallel sum reduction. As a result, varying the number of work-group sizes which is an essential parameter of parallel processing provided dissimilar computing times. Also, running at a higher sampling rate yielded a higher benchmark ratio between sequential and parallel. However, system optimization is still necessary when processing BCI in real time.