This is a straightforward port of Neural Network Ambient Occlusion (NNAO) originally written by Daniel Holden et al.
Please note that this implementation is just for initial experimental purposes. Several important things are missing, so that it's not suitable for practical use.
Due to the lack of noise reduction filter, it outputs quite noisy results. This can be improved by implementing a proper filter.
This implementation hasn't been optimized yet. It's not recommended for benchmarking or performance analysis.