Brice Videau d’Argonne prepares HIP applications
Brice Videau, a computer scientist at the U.S. Department of Energy’s (DOE) Argonne National Laboratory with years of experience in code generation, optimization, and self-tuning high-performance computing, is leading efforts to maximize the portability of HIP applications being developed for the next generation of advanced computing systems, including Argonne’s future exascale Aurora system.
The Principal Investigator of the “HIP on Aurora” project (also known as HIPLZ, or “HIP on Level Zero”), Videau holds a cross-appointment with the Argonne Leadership Computing Facility (ALCF) and Argonne’s Division of Mathematics and Computing, where his research focuses on the programming of heterogeneous platforms and self-tuning. The ALCF is a user facility of the DOE Office of Science.
HIP on Aurora
Already established as the default programming model for AMD GPUs, many important applications are expected to adopt HIP as their programming model as they target deployment on upcoming exascale systems. Additionally, HIP supports NVIDIA platforms, allowing portability of HIP applications to AMD and NVIDIA hardware. Since HIP is a likely programming model for many exascale codes, Videau’s team wants to ensure that HIP applications are as portable as possible.
The HIPLZ project, to that end, aims to investigate the best way to enable native support for HIP applications on Aurora, an Intel GPU-powered exascale system that is expected to arrive at the ALCF in 2022. Videau’s team finally leverages Level Zero, a new API that Intel is developing for its GPUs, including the Ponte Vecchio accelerators driving Aurora.
As part of the team’s development strategy, Videau and his colleagues are building on the work of a previous project that developed HIPCL, a library that allows HIP to run on OpenCL. Since Aurora will feature Level Zero as the primary low-level GPU interface, HIPLZ is redirecting efforts from HIPCL to facilitate support for HIP applications on Level Zero.
Through collaboration with other user facilities from the DOE Office of Science, Georgia Tech, and industry partners, Videau is working with an AMD team to refactor HIP so that it can support several new external backends; a first level zero backend has been implemented.
A number of application benchmarks have already been established and various proxy applications are running, demonstrating the viability of performance-critical computations. Other work in progress includes improving the functionality of frameworks, proxy apps, and real apps.
Videau’s team is currently working on developing a test suite that will allow them to evaluate the functionality and performance of HIPLZ. Of particular interest is validating its usefulness in Exascale Computing Project (ECP) applications, following successful engagement with codes such as EQSim, GAMESS, CoMeT, and NAMD.
Source: Nils Heinonen, ALCF