Matt has implemented some optimizations to the formation of Jacobian matrix on the CPU. We need to update the CUDA implementation accordingly.