SBIR/STTR Award attributes
There is a need to support most of the 100+ organizations developing new Machine Learning accelerators with software tools that will enable them to keep up with the pace at which new models and model variants are created, in more and more application domains. The tools must allow these organizations to optimize a large proportion of the Machine Learning codes, so as to focus human effort on a manageable set of high-impact code sections. Computational Science, such as deployed at the Department of Energy, is the scene of revolutionary new Machine Learning-based techniques that provide dramatically higher performance than traditional techniques. To reach their full potential, the codes implementing these techniques must be optimized. In this Phase IIB proposal, we propose to leverage work done with the R-Stream polyhedral mapper in Phases I and II to enable an effective transition of this technology into commercial Machine Learning Optimization Frameworks. The work includes new optimizations and techniques to embed the polyhedral technology into existing tools. The Phase I and II work included interoperability between the R-Stream polyhedral mapper and the LLVM (Low-Level Virtual Machine) compiler, methods to harden the R-Stream compiler, developments and optimizations of R-Stream’s TensorFlow front-end and graph optimizer, and the ability to ingest parallelism information about the input program using the OpenMP parallel programming language. The overall approach is to enable integration of the polyhedral mapper technology offered by R-Stream with newer high-level compiler toolchains, including the ones that target LLVM, with a focus on the Machine Learning software market, to which our polyhedral mapping technology is particularly relevant. We discuss existing approaches and the particular role which R-Stream and the polyhedral technology play in Machine Learning Acceleration, as well as how and why we plan to make R-Stream a central tool to the Machine Learning Acceleration community. The application of R-Stream would have a tremendous impact on the U.S. Machine Learning Acceleration effort, and on computational science capabilities.
