OpenCL Superpixel Implementation on a General Purpose Multi-core CPU
Multi-, many-core, hybrid processors and parallel programming languages are slowly becoming pervasive in main- stream computing. It is expected that they will affect a large spectrum of systems, from embedded and general-purpose, to high-end computing systems. This architectural change has already challenged programmers to efficiently write an application code that can scale over many cores to utilize its computational power. Moreover, many heterogeneous architectures exist today, hence there was an emergent need for a uniform interface to these architectures. Recently, Khronos Group defined the Open Computing Language (OpenCL) for abstracting the underlying hardware, which enables software developers to write a portable code across different shared-memory architectures. In this paper, we introduce a new parallel implementation of one of the fastest image segmentation algorithms known as Simple Linear Iterative Clustering based on OpenCL. We evaluate the effectiveness of this implementation using only multi-core GPCPU. Our implementation is fully compatible with sequential implementation. When the algorithm is executed sequentially it utilizes only 25% of total computational power of a GPCPU for any image resolution, while its modified algorithm is able to utilize close to 100% for high resolution images. The resulting algorithm is up to 5×faster than its sequential counterpart.