Running Tensorflow alongside other CUDA-code


Is it possible to make use of Tensorflow alongside other CUDA kernels (all running in the same process)?

In our use-case we wish to perform real-time decoding of video frames, manipulate the content using CUDA-kernels and finally run analysis on the output using Tensorflow.

Any ideas on how to make Tensorflow and our CUDA code play along nicely? Is it maybe possible to accomplish using clever context handling?

Thanks in advance for any help and pointers!