Implementation Analysis of Binaural Audio Crosstalk Cancellation on Heterogeneous Parallel Computing platforms using Mixed Non-Uniform Partitioned Convolution

As general DSP processor architectures don’t have massive parallel architecture, they are not suitable to implement 3D audio virtual techniques at very long filters due to computational problems. To address these implementation issues of very long filters, an efficient method called Mixed Non-uniform Partitioned Convolution is proposed in this paper for implementing binaural audio crosstalk cancellation on heterogeneous parallel computing platforms. By using massive parallel architecture of heterogeneous platforms, the proposed approach is able to solve computational problems even at filter lengths of 65536 (32-bit floating point). The partitioning scheme followed in this paper is explained in detail to schedule partitions on various compute units of GPU device. The proposed approach was implemented on AMD GPUs using task parallel concept. The instruction level optimization was also provided for complex frequency multiplication and addition using OpenCL. The performance of this approach is compared against the existing techniques proposed by Garcia and Gardener. The cost vs. computational performance tradeoff comparison was given between proposed approach and existing methods. The comparison clearly shows that proposed approach is very efficient at very long filters and requires reasonable cost of implementation in terms of number of compute units. The combination of instruction level and algorithmic level optimizations make the proposed approach more suitable for implementation of not only stereo inputs based audio CTC but also multichannel inputs, particularly at very long filter lengths on parallel computing platforms.