SRC26. GPU-Accelerated Jacobi-Like Algorithm for Eigendecomposition of General Complex Matrices
Student: Basileal Y. Imana (Trinity College)
Supervisor: Peter Yoon (Trinity College)
Abstract: Jacobi-like methods, though introduced in 1846, were not popular as they were computationally expensive for sequential computing. With the advent of parallel computing, however, it has become feasible to efficiently implement such algorithms in parallel. In addition, the Jacobi method has been shown to compute very small eigenvalues with high accuracy compared to the conventional methods. In this research, we present a novel parallel implementation of Jacobi method for eigendecomposition of general complex matrices on the GPU. Our preliminary results show a significant improvement over those on the CPU, running up to 94 times faster for general dense complex matrices of moderate size.
Two-page extended abstract: pdf