This NSF-funded project is a collaborative effort with Prof. Ali Jannesari at Iowa State University. Given the anticipated growth in HPC hardware configurations, and the long lead time for creating compilers for new architectures, it is critical to develop technology that will enable compilers to evolve rapidly for new target platforms. Since many of the translation problems in compiling are solved heuristically, it is a field that is ripe for AI. Moreover, there have been promising results in the application of Machine Learning (ML) to handle individual translation and optimization problems in compiling.

This project will enable us to learn to apply new ML algorithms, especially Graph Neural Networks, to more systematically enable the parallelization and optimization of code for heterogeneous architectures, and thereby to demonstrate how compilers and runtimes can exploit AI to autotune applications. In addition to this work, the project explores the use of compiler techniques to improve tensor program generation by learning the joint neural network and hardware features space, facilitating knowledge transfer to new, unseen target hardware.