3 ms·

Would it be a completely stupid idea to write a CUDA-based OpenCL back-end? i.e., an OpenCL-to-CUDA translator, so you can program your kernels in one single l

Would it be a completely stupid idea to write a CUDA-based OpenCL back-end? i.e., an OpenCL-to-CUDA translator, so you can program your kernels in one single language but still get the benefit of the NVidia CUDA compiler? Or are their machine models so different that that is an unreasonable thing to even try..

jjn2009 10y ago
It's likely better to just write/use libs which abstract the CUDA/OpenCL level away for generic GPU type tasks. Unless you have really strict performance requirements or are writing a very specific piece of GPU code, in which case I cant image OpenCL-to-CUDA translator would handle that case well either. They both have very similar ideas for the core of the APIs then CUDA adds a bunch of stuff on top of it for ease of use and performance which is tightly coupled with the hardware in some cases. It would be a lowest common denominator situation with a lot of those features in CUDA.