Performance Study of LU Decomposition on the Programmable GPU