Experts Chat at the oneAPI DevSummit
On December 6, 2023, a panel of experts and software development leaders met as part of the oneAPI DevSummit for AI and HPC 2023 to talk about their path towards adopting the open cross-platform SYCL* abstraction layer and the vision for SYCL’s future. We set out to discuss:
Benefits of SYCL.
Challenges in migrating from CUDA and other vendors-specific accelerated parallel programming frameworks.
Current developer pain points.
The future of SYCL and the Unified Acceleration Foundation.
All of them are deeply involved in the adoption and outreach for oneAPI, SYCL, and the newly formed Unified Acceleration Foundation (UXL) project sponsored by the Linux Foundation.
The Panelists
Dr. Steve Petruzza is an Assistant Professor at Utah State University and a Research Associate at the Scientific Computing and Imaging Institute of the University of Utah. His research spans the area of high-performance computing and large-scale scientific visualization with a focus on efficient and portable workflows for parallel and distributed data analysis and visualization, including the work on NSF National Ecology Observatory Network (NEON) survey data.
Dr. Yasaman Ghadar is a computational scientist at Argonne Leadership Computing Facility. Yasaman has a PhD in computational Chemistry and is a main Argonne collaborator for the EXAALT project, an Exascale Computing Project (ECP) effort using molecular dynamics to simulate materials relevant to nuclear fusion and fission. She also specialized in creating training materials for current and upcoming HPC machines for our user community, such as Aurora. She has organized and chaired over twenty technical workshops and forty webinar series, including seven external events in collaboration with both industrial partners and other national laboratories.
Dr. Andreas Goetz is an associate research scientist at the San Diego Supercomputer Center (SDSC) at the University of California San Diego, where he leads a research group in Data-Driven and High-Performance Computational Chemistry, working at the intersection of chemistry, life sciences, and scientific computing. He is a contributing author to the ADF and QUICK quantum chemistry software and the AMBER software for biomolecular simulations, which are widely used in academic and industrial research. Andreas also enjoys training the next generation of scientists in software engineering and numerical simulation methods.
Dr. Hugh Bird is a software engineer at Codeplay Software. He obtained his Ph.D. on “Low-order methods for the unsteady aerodynamics of finite wings” at the University of Glasgow before moving to Codeplay to work on the ComputeCpp SYCL runtime. He now works in the Performance Libraries team, contributing to the oneMKL interfaces library and DFTs in particular.
View and Listen to the full panel discussion here: |
In our panel on the “Journey Towards Ubiquitous SYCL Adoption,” we have four experts covering the benefits of SYCL, challenges in migrating from CUDA and other vendors-specific accelerated parallel programming frameworks, current developer pain points as well as the promise and future of SYCL.
Below or some of the key points they made.
Unified Acceleration Foundation and the Future of SYCL
Hugh Bird started the discussion with a summary of the Unified Acceleration Foundation’s creation and its impact on the evolution of the oneAPI initiative and SYCL adoption.
The foundation’s steering members (ARM*, Fujitsu*, Google* Cloud, Imagination Technologies*, VMware*, Qualcomm*, Samsung*, Intel) share the vision of creating a level playing field where a proprietary software stack is not the limiting factor.
Instead, we have performance portability and a common software stack that allows applications to run heterogeneously on the best hardware configuration for the job.
This is achieved through building a large open ecosystem with multivendor and multiarchitecture software framework support. An open governance model with multiple industry stakeholders and backed by the Linux Foundation* implies that it will be based on open consensus-based specifications. Hardware and software vendors contribute to it, ensuring it will not become another vendor-locked walled garden.
Key Takeaways of our Discussion
Andreas Goetz and Steve Petruzza stressed the benefit of having an open standard with cross-architecture performance portability, a growing application and algorithm ecosystem, and the ability to use the same codebase on different hardware configurations and the resource savings that come with that.
Now, with SYCLomatic and the Intel® DCP++ Compatibility Tool, the initial migration from proprietary SYCL code becomes faster and more convenient, including the handling of type mismatches between software APIs or various performance library APIs like the oneMKL interfaces project in compliance with the oneAPI specification. The latter part is especially important for users wanting to port more complex applications but with less familiarity with oneAPI and its spec elements.
Argonne Labs' adoption of oneAPI for Aurora and the scientific and educational work driven by researchers like Yasaman Gadhar provided a key proving ground to challenge and grow SYCLs and the oneAPI initiatives' readiness for wide adoption. Developments can rely on a wealth of experience, reference implementations, and workshop recordings these days.
One additional benefit of the openness of the SYCL programming framework is that you can develop an application on a relatively small platform (e.g., a laptop with a built-in GPU), including doing basic functionality testing, and then take the same code base and scale for an HPC or even supercomputer setup.
Many more things are happening, like the ongoing improvement in migration guides, in-depth documentation, tips, and tricks for running SYCL-based applications on different CPU, GPU, and accelerator combinations, as well as ongoing progress on making full performance portability a reality.
Some of the work currently being done is to provide SYCL open backend implementations for oneAPI library spec elements, including read-to-use translation code for use with runtimes targeting Intel, AMD*, NVIDIA* and other computational units.
Check It Out
To learn more about the panelists' detailed take on the state and future of SYCL, check out the full oneAPI DevSummit “The Journey to Ubiquitous SYCL Adoption” panel recording.
Get started with your own project by visiting one of the following:
More related oneAPI DevSummit content:
- SYCL Support for Continental-Scale Ecological Observations: Scalable and Portable Blending of Massive Image Mosaics
- Performant Portable HPC Applications with SYCL and oneMKL Interface
- Development and Optimization of a SYCL Backend for libCEED
Rob enables developers to streamline programming efforts across multiarchitecture compute devices for high performance applications taking advantage of Intel's family of development tools. He has extensive 20+ years of experience in technical consulting, software architecture and platform engineering working in IoT, edge, embedded software and hardware developer enabling.