Parallel science code translation stops retrying blindly: agents route each failure to a matching repair skill, beating LASSI in all 18 model-and-direction cells
Synopsis
Translating parallel scientific code among CUDA, OpenMP and SYCL has relied on fixed sequential pipelines that apply one repair strategy to every error and judge correctness only by compilation and execution. A verdict-driven orchestrator splits translation, compilation diagnosis and execution validation into three agents and routes each failure to a patch, semantic, sync or memory repair skill. On 30 HeCBench kernels, three open-source models and six translation directions, the workflow reaches a higher verified correctness rate than LASSI in all 18 model-and-direction combinations while using up to 52.4% fewer tokens.
Interpretation
Parallel scientific code translation moves from one fixed sequential pipeline to a verdict-driven orchestrator coordinating three agents, one each for translation, compilation diagnosis and execution validation, so failures are no longer retried uniformly. Earlier LLM translation pipelines applied the same repair strategy to every error regardless of root cause, which produced unnecessary model calls. Across 30 HeCBench kernels, six directions among CUDA, OpenMP and SYCL, and three open-source models, the workflow's verified correctness rate exceeds LASSI in all 18 combinations, with per-model averages of 78.6% to 94.6% against 34.5% to 65.5%.
Each failure is first classified into a specific verdict and then handed to the matching repair skill: compilation errors go to patch or memory repair, runtime crashes to sync repair, and output mismatches to semantic repair. Repair skills are dynamically loaded units, so adding a source-target language pair requires only pair-specific skill implementations and prompt templates, without changing the orchestrator or agent code. Token consumption is 29.5% to 52.4% below LASSI depending on the model, staying between 1018k and 1115k per direction; only two of the 18 cells spend more than LASSI, and both buy correctness gains of up to 60.0%.
Perspective
The demonstrated scope is 30 selected kernel-level HeCBench benchmarks translated bidirectionally among CUDA, OpenMP and SYCL; six kernels lack OpenMP implementations, leaving 24 benchmarks for OpenMP-source directions, with compilation capped at 300 seconds, execution at 120 seconds and at most 9 repair rounds. The authors propose extending the framework to multi-file codebases and full applications, and plan to bring performance feedback into the repair loop.
In-loop compilation and runtime checks overestimate the success of both systems; the offline LLM verification lowers pass rates by 8.3% to 17.3% and identifies 117 false positives, led by 31 placeholder implementations, 28 numerical errors and 25 skipped verifications. The authors also note that repair gains diminish beyond roughly five rounds and that failures surviving many rounds are systematic; the offline verifier itself can be strengthened with refined rubrics and multi-verifier voting.