Optimizing OpenCL Kernels for Iterative Statistical Applications on GPUs