Quantifying the Efficiency of High-Level Synthesis for Machine Learning Inference