high performance computing on graphics processing units: hgpu.org

For the blind separation of convolutive mixtures, a huge processing power is required. In this paper we propose a massive parallel implementation of the Independent Component Analysis in the time-frequency domain using the processing power of the current graphics adapters within the CUDA framework. The often used approach for solving the separation task is the transformation to the time-frequency domain where the convolution becomes a multiplication. This allows for the use of an instantaneous ICA algorithm independently in each frequency bin, which greatly reduces complexity. Besides algorithmic simplification, this approach also provides a very founded approach for parallelization. In this work, we propose an implementation using the CUDA framework, which provides an easy interface for the implementation of massive parallel algorithms. The new implementation allows for a speedup in the order of two magnitudes, as it will be shown on real-world examples.