Speaker
Description
Today’s supercomputers are massively parallel, heterogeneous systems and GPUs have long been established as accelerators to boost the execution of data parallel algorithms. However, programming GPUs remains complex and requires specialized expertise to optimize code for a specific architecture, oftentimes involving a vendor-specific programming model. This usually results in code that is not portable between GPU vendors, typically requiring that programmers have to maintain multiple versions to target different architectures. Many software packages reflect this circumstance and provide tailored implementations for HIP, CUDA, and OpenCL. This talk has two parts. Firstly, it will introduce target offloading using the OpenMP API, an easy and portable way to exploit a GPU's parallelism. We will review some basic OpenMP directives and allude to key concepts of the heterogeneous features of the OpenMP API. Secondly, the talk will highlight some recent case studies of (large) Fortran codes that have been ported to AMD Instinct™ Accelerators using the AMD Next-Generation Fortran Compiler, based on LLVM/Flang. For the case studies, we will sketch the way the OpenMP API was applied and review the performance obtained on AMD Instinct™ Accelerators.
