On the Portability of GPU-Accelerated Applications via Automated Source-to-Source Translation