by dar » Sat Jan 11, 2014 5:18 am
Following up on the discussion of OpenCL wrappers. I would avoid OpenCL wrappers. If the OpenCL API cannot stand on its own, wrappers are not a long term solution. A significant issue is that Epiphany is not a GPU and OpenCL was designed around a GPU. So OpenCL GPU code will take you astray. We have learned this from experience.
Consider STDCL. Its not a wrapper but it leverages OpenCL for portability, at the moment at least, the way other APIs do, like OpenACC (though its very different from OpenACC). STDCL supports Parallella now with the COPRTHR SDK v1.6 released yesterday. The SDK also provides a more direct API that may be of interest in the context of getting OpenMP to work on Parallella for any project aimed in that direction. Many OpenMP implementations use pthreads for a thread pool. The COPRTHR direct API includes a logical extension of pthreads to co-processors. Parallella is a target for support. That may provide a useful connection in the future.
So there are really two interwoven issues here. If you want a quick example of parallel code for Parallella, STDCL is very simple and there are a lot of tools in place, like clcc that supports a standard compilation model - these things are available now, not future projects.
If you are looking for a programming quick start with examples for Parallella, we are hoping to post something within days. No tedious verbose OpenCL headaches. Just simple clean code.
