Extracting features for HLS kernels using CodeBERT

This repo uses CodeBERT, a multi-programming-lingual model pre-trained on NL-PL pairs in 6 programming languages, in order to extract features from HLS kernels. The features are then visualized with dimensionality reduction methods and clustered in order to gain insights on the similarity of their execution.

Motivation and results

See the presentation in NLP_for_HLS.pdf.

Data

See dataset folder.

Dependencies

See requirements.txt. Optional: CUDA support for torch.