A recent article in the Journal of Parallel Computing looks at the challenges faced in creating a robust CUDA-to-OpenCL translator, presents an instantiation of a CUDA-to-OpenCL (CU2CL) source-to-source translator, and evaluates its efficacy on real CUDA codes. The researchers conclude that although it is not straightforward and (currently) subject to some important limitations, robust automatic source translation from CUDA to OpenCL is largely achievable. Further they show that once translated, when executed on the same device, application performance is
retained, suggesting that the improved portability of OpenCL codes no longer results in reduced performance on CUDA devices. Finally, they present a robust automatic translator capable of reducing the man-weeks required for manual translations
to the order of seconds.
Read the full paper at: http://synergy.cs.vt.edu/pubs/papers/gardner-cu2cl-parco13.pdf