IASCI Research Publishing
Journal of Advanced Discovery and Artificial Intelligence
Research Article Published October 9, 2026 Article jadai20260008

Agent-Guided CUDA Kernel Generation for High-Performance Numerical Computing

Read & download PDF
Abstract

High-performance numerical computing requires code that is not only functionally correct but also optimized for parallel execution, memory access, and hardware utilization. Current large language models can generate CUDA kernels, but they often produce inefficient thread mapping, poor memory coalescing, excessive synchronization, and unstable performance across GPU architectures. This study examines agent-guided CUDA kernel generation for numerical computing tasks. We propose KernelAgent, a multi-agent model composed of a mathematical decomposition agent, a CUDA code generation agent, a correctness testing agent, a profiling agent, and a performance repair agent. The mathematical decomposition agent converts numerical tasks into parallelizable operations. The code generation agent produces CUDA C++ kernels and host-side launch code. The correctness testing agent compares GPU outputs with CPU reference implementations under random and boundary inputs. The profiling agent collects occupancy, memory throughput, warp divergence, register usage, and kernel execution time. The performance repair agent rewrites kernels using shared memory tiling, loop unrolling, vectorized memory access, and block-size adjustment. The model was evaluated on 860 numerical programming tasks, including matrix multiplication, convolution, prefix sum, stencil computation, reduction, sparse operations, and element-wise tensor transformations. Experiments generated 10,320 CUDA programs and ran 64,800 correctness and profiling tests on NVIDIA RTX 4090, A100, and L40S GPUs. Compared with a single-agent code generator, KernelAgent improved functional correctness from 62.1% to 80.4%. Average kernel execution time decreased by 36.7%, global memory load efficiency increased from 68.5% to 84.9%, and achieved occupancy increased from 51.3% to 69.8%. For matrix and convolution tasks, KernelAgent achieved a 1.72× average speedup over unoptimized generated kernels. These findings indicate that agent-level profiling and repair can substantially improve the quality of generated GPU programs.

Keywords
CUDA code generationnumerical computingGPU optimizationperformance profilingmulti-agent systemsparallel programmingkernel repair
References
  1. Chen, F., Liang, H., Li, S., Yue, L., & Xu, P. (2025). Design of Domestic Chip Scheduling Architecture for Smart Grid Based on Edge Collaboration.
  2. Que, H. H., Jin, Y., Wang, T., Liu, M. K., Yang, X. H., & Qiao, F. (2023). A Survey of Approximate Computing: From Arithmetic Units Design to High-Level Applications. Journal of Computer Science and Technology, 38(2), 251-272.
  3. Liang, R., Feifan, F. N. U., Liang, Y., & Ye, Z. (2025). Emotion-Aware Interface Adaptation in Mobile Applications Based on Color Psychology and Multimodal User State Recognition. Frontiers in Artificial Intelligence Research, 2(1), 51-57.
  4. Zhang, Z. (2026). A Study on the Impact of Cost Presentation on Decision-making Bias and Cancellation Behavior in Online Subscriptions. Available at SSRN 7349839.
  5. Fan, Y., Miyazaki, T., Tang, Z., Wang, J., Huang, Y., & Omachi, S. (2025, November). Scene Text Reconstructor: A Contextual-Aware Masking Framework for Pre-training Text Detectors. In International Conference on Neural Information Processing (pp. 181-195). Singapore: Springer Nature Singapore.
  6. Zheng, J., & Makar, M. (2022). Causally motivated multi-shortcut identification and removal. Advances in Neural Information Processing Systems, 35, 12800-12812.
  7. Qi, C., & Qiao, X. (2026). Building Reliable AI Systems: A Framework That Bridges Architecture and Operations. Available at SSRN 6795298.
  8. Xiong, W., Zeng, Y., & Guo, Y. (2026). A Study on the Impact of MEP Space Organization in Large-Scale Commercial Complexes on Investment Efficiency and Its Mechanisms.
  9. Mini Han Wang , " AI-Powered Innovations in Ophthalmic Diagnosis and Treatment ", Bentham Science Publishers (2025). https://doi.org/10.2174/97988988122871250101
  10. Hong, Z., Liang, S., Xu, T., & Chen, H. (2026). A Study on Failover Verification and Recovery Objective Prediction for Cross-Region Cloud Services.
  11. Huang, J., Xu, T., Yang, J., & Yin, J. (2026). A Graph Neural Network Approach for Financial Data Validation and Risk Identification in Large-Scale Intercompany Transactions. Available at SSRN 7175058.
  12. Zhao, J., Fan, J., & Li, L. (2026). A Study on an Explainable Causal-Enhanced LLM Agent for Predicting the Forming Quality of Automotive Component Materials.
  13. Liu, W., Zhao, R., Sha, Z., Cui, Q., & Zhang, Y. (2026). Mapping the City Through the Lens of Language Models. arXiv preprint arXiv:2608.02971.
  14. You, S. (2026). Verifiable Audit Mechanisms in AI Compliance Automation: Scalability. Available at SSRN 6547458.
  15. Jiao, Y., Shi, T., Zhao, B., & Wang, A. (2026). Retrieval-Guided Structured Reasoning and Interpretable Representation Learning for Large-Scale Video–Language Models.
  16. Liang, S., Du, Y., Chen, W., & Liu, Z. (2026). Order Allocation and Emergency Transportation Decisions Amid Multi-Source Procurement Disruptions.
  17. Zhao, R., Liu, W., Sha, Z., Su, N., Zhang, Y., & Long, Y. (2026). Culturally uneven urban perception in large language models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2604.20048v3.
  18. Yuan, Y., Huang, J., Yin, J., & Xu, T. (2026). Confidence Calibration and Semantic Error Analysis for Ambiguous Coreference Resolution in User-generated Social Media Text. Available at SSRN 7393238.
  19. Zhang, G., Xu, Y., Zhang, M., Wang, S., & Lin, N. (2021). The brain network in support of social semantic accumulation. Social cognitive and affective neuroscience, 16(4), 393-405.
  20. Bao, Y., & Wang, H. (2026). Large-scale Metrics Migration: A Systematic Approach to Rebuilding Production Measurement Infrastructure. Available at SSRN 7185660.
  21. Han, Z., Chen, W., Han, Y., Mao, R., & Qin, J. (2026). Fast Diversified Top-k Rule Discovery via User-Guided Embeddings. IEEE Transactions on Knowledge and Data Engineering.
  22. Bao, Y., & Wang, H. (2026). Orion: A High-Throughput, Fault-Tolerant Event Routing Architecture for Hyperscale Data Streams. Fault-Tolerant Event Routing Architecture for Hyperscale Data Streams (January 01, 2026).
  23. Wang, S., Feng, Y., & Fang, X. (2026, May). A Large Language Model-Enabled Multi-Agent Collaboration Method for Complex Task Solving. In 2026 6th International Symposium on Computer Technology and Information Science (ISCTIS) (pp. 253-256). IEEE.
  24. Chaoyong Ma, Nan Si, Xuewei Song, Chen Liang and Haoming Luo, A Bearing Fault Diagnosis Method Based on Concise Empirical Wavelet Transform, International Journal of Comprehensive Engineering，2023, 12(1), 1-6.
  25. Yuyan Li, Lin Chen, Mengyu Yang, Qiaomei Li and Taifu Li, Polarization Converters Composed of Metasurfaces with Independent Manipulation of Polarization and Phase, International Journal of Comprehensive Engineering，2023, 12(1), 7-14.
  26. Zhao, Z., & Welsch, R. E. (2024). Hierarchical reinforced trader (hrt): A bi-level approach for optimizing stock selection and execution. arXiv preprint arXiv:2410.14927.
  27. Jiaxun Du, Zhenbao Fu, Jiachen Liu, Huaqing Wang, Zhentao Ge and Liuyang Song, Bearing Fault Diagnosis Method Based on Attention Residual Network Under Small Sample Condition, International Journal of Comprehensive Engineering，2023, 12(1), 15-24.
  28. Yang, Z., Alexandrova, A. N., & Sautet, P. (2026). Modeling CO2 Hydrogenation to Methanol on an Ensemble of Inverse ZrO2 on Cu Catalytic Sites: Mechanism, Reactivity, and Deactivation. Angewandte Chemie, e5448247.
  29. Jiawei Lin, Changkun Han, Liuyang Song, Wei Lu and Huaqing Wang, Time Domain Feature Enhancement Model Based on Correlated B-spline Wavelet Base Convolutional Sparse, International Journal of Comprehensive Engineering，2023, 12(1), 25-33.
  30. Zhao, Z., & Welsch, R. E. (2026). Point-in-Time Financial RAG with Frozen LLMs and Market-Feedback Adaptive Retrieval. arXiv preprint arXiv:2605.31201.
  31. Quan, Q., Gao, Y., & Wang, Q. (2025). The impact of teacher emotional support on students' engagement in AI-mediated English learning environments: The mediating role of resilience and self-efficacy. Acta Psychologica, 260, 105766.
  32. Wang, Z., Yu, J., Liu, H., Zheng, Z., Jin, Y., Li, S., ... & Zhang, K. (2025, July). Generative Music Models’ Alignment with Professional and Amateur Users’ Expectations. In Findings of the Association for Computational Linguistics: ACL 2025 (pp. 6909-6920).
  33. Xu, Y., Sun, Y., Zhai, B., Li, M., Liang, W., Li, Y., & Du, S. (2025, April). Zero-shot video moment retrieval via off-the-shelf multimodal large language models. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 39, No. 9, pp. 8978-8986).
  34. Zheng, X., Chen, X., Gong, S., Griffin, X., & Slabaugh, G. (2025, September). Xfmamba: Cross-fusion mamba for multi-view medical image classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 672-682). Cham: Springer Nature Switzerland.
  35. Su, J., Cai, C., Zhu, F., He, C., Xu, X., Guan, D., & Si, C. (2024, September). Momentum auxiliary network for supervised local learning. In European Conference on Computer Vision (pp. 276-292). Cham: Springer Nature Switzerland.
  36. Su, J., Zhu, F., Shi, H., Han, T., Qiu, Y., Luo, J., ... & Gao, J. (2025). MAN++: Scaling Momentum Auxiliary Network for Supervised Local Learning in Vision Tasks. arXiv preprint arXiv:2507.16279.
  37. Li, P., Zhang, H., Li, W., Huang, D., Guo, Z., Chen, J., ... & Koshizuka, N. (2025). GeoAvatar: A big mobile phone positioning data-driven method for individualized pseudo personal mobility data generation. Computers, Environment and Urban Systems, 119, 102252.
  38. Lyu, Y., Liu, K., Cao, Z., Luo, Y., Yin, L., Yang, T., & Chen, Z. (2026). FLAIR: An LLM-integrated multi-layer framework for fine-grained activity type inference in human trajectory data. Transportation Research Part C: Emerging Technologies, 190, 105787.
  39. X. Chen, C. Huang, G. Chen, D. Feng and P. Xiao, "The Communication and Computation Trade-Off in Wireless Semantic Communications," in IEEE Wireless Communications Letters, vol. 14, no. 7, pp. 2259-2263, July 2025, doi: 10.1109/LWC.2025.3569205.
  40. Cui, J., Zhang, G., Chen, Z., & Yu, N. (2022). Multi-homed abnormal behavior detection algorithm based on fuzzy particle swarm cluster in user and entity behavior analytics. Scientific Reports, 12(1), 22349.
  41. Wang, T., & Xia, Z. (2025). Stability of In-Context Learning: A Spectral Coverage Perspective. arXiv preprint arXiv:2509.20677.
Publication details
Journal
Journal of Advanced Discovery and Artificial Intelligence
Volume
1 (2026)
Issue
1 · Forthcoming issue
Article number
jadai20260008
License
CC BY 4.0