What does this research mean for the field?

Extending the Oxford Parallel Structured mesh solver library (OPS) with reverse-mode algorithmic differentiation enables the generation of performance-portable adjoints for structured mesh applications on CPUs and GPUs with computational overheads comparable to state-of-the-art tools. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To enable performance-portable, adjoint-mode algorithmic differentiation within the Oxford Parallel Structured mesh library across diverse hardware architectures.

September 10, 2026ACM Transactions on Mathematical Software

Performance portable adjoints for structured mesh applications with OPS

View Full Paper

Ask AI

Bookmark

Authors

GBGábor Dániel Balogh JLJohannes Lotz JTJacques du Toit

Discussion

Member takes

Overview

Framework evaluation demonstrates performance-portable adjoint differentiation across structured mesh codes, indicating efficient CPU and GPU gradient computation for scientific optimization.

Key Points

To enable performance-portable, adjoint-mode algorithmic differentiation within the Oxford Parallel Structured mesh library across diverse hardware architectures.
Extended the Oxford Parallel Structured (OPS) domain-specific language to support reverse-mode algorithmic differentiation via loop-level adjoint taping.
Generated platform-specific parallel implementations targeting OpenMP on CPUs and CUDA on GPUs from user-supplied primal and adjoint stencil kernels.
Benchmarked the framework against three structured mesh applications to assess execution overhead and cross-platform performance.
Derivative calculations required 3.7x to 9.7x the runtime of the original applications, inclusive of the primal evaluation step.
Achieved performance portability across both CPU and GPU platforms, matching the efficiency standards of state-of-the-art differentiation tools.

Cite This Study

Balogh et al. (2026) studied this question.

synapsesocial.com/papers/6aa27bbd58559d80afc7524b https://doi.org/10.1145/3815776

View Full Paper

Ask AI

Bookmark

September 10, 2026ACM Transactions on Mathematical Software

Performance portable adjoints for structured mesh applications with OPS

View Full Paper

Ask AI

Bookmark

Authors

GBGábor Dániel Balogh JLJohannes Lotz JTJacques du Toit

Discussion

Member takes

Overview

Framework evaluation demonstrates performance-portable adjoint differentiation across structured mesh codes, indicating efficient CPU and GPU gradient computation for scientific optimization.

Key Points

To enable performance-portable, adjoint-mode algorithmic differentiation within the Oxford Parallel Structured mesh library across diverse hardware architectures.
Extended the Oxford Parallel Structured (OPS) domain-specific language to support reverse-mode algorithmic differentiation via loop-level adjoint taping.
Generated platform-specific parallel implementations targeting OpenMP on CPUs and CUDA on GPUs from user-supplied primal and adjoint stencil kernels.
Benchmarked the framework against three structured mesh applications to assess execution overhead and cross-platform performance.
Derivative calculations required 3.7x to 9.7x the runtime of the original applications, inclusive of the primal evaluation step.
Achieved performance portability across both CPU and GPU platforms, matching the efficiency standards of state-of-the-art differentiation tools.

Cite This Study

Balogh et al. (2026) studied this question.

synapsesocial.com/papers/6aa27bbd58559d80afc7524b https://doi.org/10.1145/3815776

View Full Paper

Ask AI

Bookmark