The computational demand for real-time control systems has significantly increased, particularly with the use of machine learning, which has been met by higher core counts on embedded architectures. Mainstream programming languages support such parallelism for general purpose programming, but they lack native support for expressing parallelism under real-time constraints. While external tools and realtime operating systems offer temporal guarantees, they often fall short in providing integrated, language-level constructs for both coarse- and fine-grained parallelism. OpenMP is a natural candidate for this role, but lacks support for real-time scheduling and synchronization under deadline constraints. This work introduces OpenMP-RT to raise a discussion for inclusion in the OpenMP standard. OpenMP-RT is a framework that extends OpenMP to support real-time periodic tasks and predictable inter-task communication. We identify the limitations of standard OpenMP for real-time applications and explore potential extensions to the OpenMP 5.1 specification to address these gaps. Our framework introduces a new task construct, rttask, and a configuration-driven specification model. OpenMP-RT enables time-predictable, lock-free communication across priority levels and supports both static-priority and EDF scheduling. We have implemented OpenMP-RT in the LLVM C compiler under OpenMP 5.1 and analyzed its performance in terms of deadlines and real-time synchronization.
OpenMP target offloading currently supports only four basic devicetodevice copy routines, a gap that forces programmers to rely on vendor libraries and limits scalability on multiGPU nodes. We introduce a prototype OpenMP-native, topology-aware collective-communication framework that: (i) exposes lightweight APIs for querying link type and peer bandwidth, (ii) offers an optional devicegroup abstraction, and (iii) implements ring and pipelinetree broadcasts whose ordering is optimized at run time. On an eightGPU MI250X node, a topologyaware halo exchange accelerates communication by up to 23.4
Quantum accelerators hold promise for applications including combinatorial optimization, chemistry simulation, and hybrid ML workloads, etc. To program such applications, most quantum SDKs (e.g., Qiskit, Cirq) expose only Python APIs. This leaves C/C++ codesmany already orchestrating heterogeneous CPU/GPU work with frameworks such as OpenMPwithout a clean path to invoke quantum kernels or iterate on their results. This work contributes OpenMP-Q and aims to raise a discussion for inclusion in the OpenMP standard. OpenMP-Q is an extension to the OpenMP 5.x target model that treats a quantum processor as a first-class device and enables bi-directional communication between C++ and Python. OpenMP-Q adds (i) a quantum device type, (ii) a mechanism that bridges OpenMPs offload API to the Python quantum stack via POSIX pipes, and (iii) an extended offloading framework that communicates variables via map clauses to a Python script, executes the quantum task, and reinserts the results into the same mapped variables. The pipe mechanism also supports reverse-offload-style workflows, e.g., variational algorithms where C++ supplies initial rotation angles, the quantum backend returns qubit statistics, and the host refines those angles for subsequent iterations, without breaking OpenMPs asynchronous execution model or requiring extra device kernels. Moreover, by combining standard OpenMP parallel regions with per-thread quantum offload, OpenMP-Q scales natively across multiple QPUs, dispatching independent circuit evaluations in parallel
Fine-tuning existing LLMs for specialized tasks has become a very attractive alternative due to its low cost and quick development cycle. With many pre-trained LLMs available, it is an increasingly complex task to choose the correct model as the starting point or base model. In this work we discuss ChatPORT - a specialized fine-tuned LLM geared towards providing correctly translated codes from one programming model to another. We evaluate a number of base models and compare and contrast their features and characteristics that make them a viable starting point. In this paper, we focus on the OpenMP offload porting capabilities of ChatPORT. We build our training data using kernels from the Heterogeneous Computing Benchmarks (HeCBench) [12] and the OpenMP Validation and Verification suite [5] to fine-tune the base models. We then test the model using unseen kernels extracted from the HeCBench benchmark suite. Our results show that: (1) not all open LLMs geared towards HPC are aware of programming models like OpenMP, (2) although all base models benefit from fine-tuning they learn differently and produce different correctness rates, (3) depending on the memory size and compute resource available, different base models can be used for fine-tuning without significantly affecting the quality of transpiled code they generate, (4) fine-tuning improved the correctness rate of the LLM by an average of 43.2
Modern deep learning and HPC applications on distributed GPU clusters rely on collective communication primitives such as Broadcast, All-Reduce, and All-Gather for efficient data coordination across devices and nodes. The current OpenMP specification lacks native support for multi-device collectives within its target offloading model, leading developers to re-implement algorithms or depend on vendor-specific libraries. Recent OpenMP extensions aim to support distributed clusters, but naive collectives in such environments often suffer from performance issues due to network topology heterogeneity. In this paper, we present the OpenMP Collective Communication Library (OMPCCL), an extension to the OpenMP Target model that integrates portable collective primitives and device group semantics into the OpenMP interface. We prototype several collectives including Broadcast, All-Reduce, and Reduce-Scatter and evaluate performance on a 64-GPU cluster. Results show up to 5× speedup over naive methods, while maintaining OpenMP’s portability and simplicity. We also analyze the influence of network topology on collective performance and provide guidelines for efficient OpenMP collectives on GPU clusters.
