We describe the design of a C++ vector-manipulation substrate that allows first-order optimization algorithms to be expressed in a concise and readable manner, yet still achieve high performance in parallel computing environments. We use standard object-oriented techniques of encapsulation and operator overloading, combined with a novel “symbolic temporaries” delayed-evaluation system that greatly reduces the overhead induced by compiler temporaries and economizes on memory references. We also provide infrastructure to support line-search methods by caching function values and gradients at previously-visited points in a transparent manner that does not “clutter” the principal implementation. We demonstrate the usefulness of our vector-substrate tools by employing them to efficiently solve large-scale LASSO problems using hundreds of processor cores. We reformulate the LASSO problem as a bound-constrained quadratic optimization, and then solve it using the Spectral Projected Gradient (SPG) method implemented through our vector-manipulation substrate.
RUTCOR, Rutgers University