https://hgpu.org/?p=1429
RapidMind: Portability across Architectures and its Limitations