[barneso]

The Tensorflow code mentions "GCUDACC" in several places, and from the surrounding comments it seems to be targeted at OpenCL as well as CUDA. So it seems that this has been at least considered.