While working on the DARPA PAPPA solicitation, Dr. Mertoguno developed the concept of software distillation to facilitate automated code parallelization improvement. This approach involves approximating and extracting the functionality of legacy code into a highly abstract representation. We leverage graph neural networks to guide this “lifting” process, converting existing code functionality into an incomplete yet sufficiently abstract form. Using the original program’s input-output behavior as a reference, we then iteratively re-synthesize replacement code based on this abstracted representation—effectively transforming old code into improved, more parallelized-able versions.
Under the DARPA V-SPELLS program, we adapted the software distillation concept to perform two distinct types of lifting: one that recognizes the code’s functionality and another that identifies its computational domain. Here, a domain-specific language (DSL) emerges naturally as the cross product of the extracted functionality and the computational domain. We further employed domain lifting as part of our proposed solution for DARPA MOCHA, arguing that processors and accelerators are typically optimized for certain computation domains. By performing domain lifting and characterizing inter-domain data dependencies, compilers can generate highly optimized executable that run efficiently across a diverse range of processors and accelerators.
