International Journal of Engineering and Technology

Partitioned convolutions are the best methods to address the system performance related issues in 3D virtualization techniques both in terms of latency and computational complexity. General DSP processor architectures are not suitable to implement very long filters due to increase in computational complexity and required on-chip memory. In this paper, an efficient method called Mixed Non-uniform partitioned convolution is explained to overcome computational problems for implementing three channel OSD (Optimal Source Distribution) with stereo inputs on heterogeneous parallel computing platforms. With the massive parallel computing architecture, the partitioning scheme used for this method prove that it is possible to implement OSD system containing 6 filters, each filter has a filter length of 65536 (32-bit floating point) on these platforms. The proposed algorithms were implemented on AMD based Bonaire GPU using task parallelism. The advantage of proposed method is that it provides zero output latency, which is desired in real-time applications. The computational performance and the system cost of proposed method was compared with existing approaches. The performance comparison clearly provides information that the proposed approach is suitable for implementation of OSD system at very long filter lengths with reasonable system cost in terms of compute units.