ideas - how to run convolutional nets on SDL2 surfaces

I’d like to be able to run a pre-trained convolutional neural net on an SDL2 surface, in a crosplatform way. The image itself might come from software rendering, or openGL. Using C++ (also interested in doing this from Rust, C-FFI bindings). These nets would do smart upscaling, and ideally other more elaborate enhancements. The use case is enhancing games (2D,3D)

What CNN library would have the highest chance of working well this way?
I’ve implemented convolutional nets myself in OpenCL many years ago, but am reluctant to go that route again because this API is out of favour, and there’s dedicated accelerators (Nvidia tensor cores and other s) which probably need some vendor supported API.

any ideas on how to do this?