Eleventh Workshop on Accelerator Programming and Directives (WACCPD 2024)

Program Chairs

Maciej Cytowski, Pawsey Supercomputing Research Centre, Australia
Emma Tolley, École Polytechnique Fédérale de Lausanne, Switzerland
Joseph Schoonover, Fluid Numerics, USA

General Chairs

Sridutt Bhalachandra, Lawrence Berkeley National Laboratory, USA
Sunita Chandrasekaran, University of Delaware, USA
Guido Juckeland, HZDR, Germany

WACCPD2023 Proceedings are now available

WACCPD 2023 conference proceedings 

SC24 Conference

WACCPD 2024 is held in conjunction with SC24, organised in Atlanta, GA, 17-22 November 2024.

Our workshop has been scheduled for: Mon Nov 18, 2024, 2pm-5:30pm

About the Workshop

The inclusion of accelerators in HPC systems is becoming well established and there are many examples of successful deployments of heterogeneous systems today. We expect this trend to continue: accelerators are becoming even more widely used, a larger fraction of the system compute capability will be delivered by accelerators, and there will be an even tighter coupling of components in a compute node. The change in the HPC system landscape, enabled by both the increasing capability and usability of accelerators such as GPUs, ML/AI chips and QPUs, opens new computational possibilities and creates challenges related to algorithm design, portability and standardization of programming models. Technology enablers look for opportunities to further integrate compute components. In the context of GPUs, these efforts include higher bandwidth memory technologies with larger capacities, hardware-managed caches, and the ability to access CPU data directly. As a result, scientific software developers are offered a rich platform to exploit the multiple levels of parallelism in their applications. This year’s workshop will look at bringing together professionals working in the field of accelerated HPC and individuals exploring new accelerator technologies, especially quantum computing. This will allow a discussion on similarities, differences, portability and standardization of programming models.

In today’s HPC environment, systems with heterogeneous node architectures providing multiple levels of parallelism are omnipresent. The next generation of systems may feature GPU-like accelerators combined with other accelerators to provide improved performance for a wider variety of application kernels. Examples include ML/AI hardware (TPUs), FPGAs and, in the future, QPUs. This would introduce further complexity to application programmers because different programming languages and frameworks (e.g CUDA and HIP) may be required for each architectural component in a compute node. This type of specialization complicates maintenance and portability to other systems. Thus, the importance of programming approaches that can provide performance, scalability, and portability while exploiting the maximum available parallelism is increasing. It is highly desirable that programmers are able to keep a single code base to help ease maintenance and avoid the need to debug and optimize multiple versions of the same code. In the context of ML/AI hardware, this includes the use of different precision arithmetic for scientific computations and the design and precision recovery. In the context of quantum computing, the standardization of computing models is extremely important, especially in the context of leveraging existing hybrid programming models and enhancing compiler frameworks to support new hardware.

Exploiting the maximum available parallelism out of such systems necessitates refactoring applications and using a programming approach that can make use of the accelerators. Historically, the favored portable approaches, and sole focus of our earlier workshops, were OpenMP offloading and OpenACC, both based on directives. Today, we recognize the evolution of other options to adapt to heterogeneity and, starting in 2021, we extended the workshop scope to include use of Fortran/C++ standard language parallelism, SYCL, DPC++, Kokkos, RAJA as well as task-based and data-centric models like Regent, Legion, OmpSs among several alternatives that can provide scalable as well as portable solutions without compromising on performance. A programmer’s expectation from the software community is to deliver solutions that would allow maintenance of a single code base whenever possible, thus avoiding duplicate effort across programming models and architectures.

Software abstraction-based programming models such as OpenMP and OpenACC have been serving this purpose over the past several years and are likely to represent one path forward. These programming models address the ‘X’ component in a hybrid MPI+X programming approach by providing programmers high-level directives and delegating some burden to the compiler. With the increased importance of other programming models to be considered in this workshop (e.g. SYCL, DPC++, Kokkos, RAJA, Regent, Legion, OmpSs), there may be other challenges and opportunities to efficiently distribute computations across multiple nodes. In the context of quantum computing, we can see attempts to achieve inter-quantum portability using well established languages like Qiskit, but also developing new portable models such as OpenQL and leveraging existing task offloading models, including OpenMP.

Our intent is to share methods and case studies demonstrating programmability, performance, and performance portability across architectures for a multitude of distributed HPC, data, AI and quantum computing workloads.

Workshop Important Deadlines

Workshop format

WACCPD is a workshop centered on peer-reviewed and published technical papers. The workshop will have one invited talk. Depending on the type of submission talks will be scheduled for 15+5min (Full Paper) and 5+5min (Extended Abstract).

Impressions from the 2023 Workshop