This supports OpenCL (Samsung needed it for TV's).

https://velesnet.ml/

However the benchmarks for OpenCL look about 5x slower than CUDA