OpenCL implementation of Cholesky matrix decomposition

This paper presents some OpenCL implementations for Cholesky decomposition, a very popular algorithm used in linear algebra and signal processing applications. The Cholesky algorithm represents a very interesting candidate for OpenCL implementation since it contains sequential parts besides parallel ones. Furthermore, one step involves just a small amount of calculations. These characteristics pose challenges which call for suitable techniques to overcome the limitations of the language. We propose several versions of the implementation of the Cholesky algorithm, then provide an analysis of the trade off between complexity and performance offered by each of them. We also analyze the differences between execution of the program on GPU and on multicore CPU.