The crucial parts of the code (precomputing the feature matrices and objective evaluations) are parallelized, which gives a significant performance boost on multi-core machines. Also, the RELEASE target is much faster (about a factor of 10) than the DEBUG target, which is presumably due to disabling debug checks/output on the preprocessor level and optimizing the code aggressively (-O3).

pulse-learning uses an L-BFGS library for optimizing the feature weights. For Linux (at least Arch and Ubuntu) there are ready-made packages available.