Towards a framework for abstracting accelerators in parallel applications: experience with cell