> So, you allocate memory in C-land and have Go structs that wrap your pointers. However, now cleanup becomes interesting. You do not want to rely on finalizers

> So, you allocate memory in C-land and have Go structs that wrap your pointers. However, now cleanup becomes interesting. You do not want to rely on finalizers, since they are not guaranteed to run. I did similar thing with C# more than once. Not with TensorFlow, but with my only C++ libraries that also used SIMD and therefore required aligned memory buffers. It worked just fine. There’s IDisposable for deterministic cleanup, and finalizers as a safety net. Unmanaged interop, i.e. [DllImport], is supported on all platforms, e.g. on Linux it imports from *.so libraries.