Using Machine Learning for OpenMP GPU Offloading in LLVM
Event Type
ACM Student Research Competition: Graduate Poster
ACM Student Research Competition: Undergraduate Poster
Posters
Student Program
TP
XO
TimeWednesday, 18 November 20208:30am - 5pm EDT
LocationPoster Module
DescriptionOpenMP 5.0 provides features to exploit the compute power within the node of today's leadership class facilities. Among these features, the GPU offloading directives are key to take advantage of heterogeneity on modern machines. These features place the domain scientists with portability challenges, however, especially for optimizing data movement between a host and a device. Tools that facilitate the usage of such features are becoming important to the scientific community. An important tool for porting legacy codes to newer machines will be compilers that can predict the feasibility of transferring kernels on GPUs and insert required OpenMP GPU offload features automatically at compile time. In this work, we are exploring a novel approach for the automated handling of OpenMP GPU offloading using machine learning techniques. We aim to develop an end-to-end application framework, from legacy code to GPU offloading, that integrates machine learning techniques into the LLVM compiler.