CUDA-Harness: Harnessing Agentic CUDA Kernel Generation and Optimization from Natural Language
CUDA-Harness: Harnessing Agentic CUDA Kernel Generation and Optimization from Natural Language
CUDA-Harness：利用自然语言实现代理式 CUDA 内核生成与优化
Developing high-performance CUDA kernels demands specialized knowledge in algorithm implementation, correctness validation, and hardware-aware parallel optimization, creating a substantial expertise barrier and making generating CUDA kernels directly from natural language (Text2CUDA) essential.
开发高性能 CUDA 内核需要算法实现、正确性验证以及硬件感知并行优化方面的专业知识，这构成了巨大的专业门槛，使得直接从自然语言生成 CUDA 内核（Text2CUDA）变得至关重要。
Meanwhile, the general-purpose code generation capability of Large Language Models (LLMs) prompts a series of works exploring LLM-based CUDA kernel generation. They mainly focus on transpilation from high-level frameworks such as PyTorch to CUDA (Torch2CUDA) rather than Text2CUDA, where models must understand the high-level input semantics and handle low-level kernel implementation and validation.
与此同时，大语言模型（LLM）的通用代码生成能力促使了一系列探索基于 LLM 的 CUDA 内核生成的研究。这些研究主要集中在从 PyTorch 等高级框架到 CUDA 的转译（Torch2CUDA），而非 Text2CUDA；在 Text2CUDA 中，模型必须理解高级输入语义，并处理底层的内核实现与验证。
Additionally, these methods are vulnerable to reward hacking due to reliance on predefined test inputs. In this paper, we propose CUDA-Harness, a framework for harnessing agentic CUDA kernel generation and optimization from natural language.
此外，由于依赖预定义的测试输入，这些方法容易受到奖励欺骗（reward hacking）的影响。在本文中，我们提出了 CUDA-Harness，这是一个用于利用自然语言进行代理式 CUDA 内核生成与优化的框架。
Specifically, we introduce Intermediate-Structured Generation to connect high-level semantic understanding with low-level kernel generation. To dilute reward hacking in Text2CUDA, we construct Synthesis-Based Verification to provide isolated test data and progressive validation.
具体而言，我们引入了“中间结构化生成”（Intermediate-Structured Generation）来连接高级语义理解与底层内核生成。为了削弱 Text2CUDA 中的奖励欺骗，我们构建了“基于合成的验证”（Synthesis-Based Verification），以提供隔离的测试数据和渐进式验证。
Furthermore, we propose Feedback-Adaptive Evolution, a kernel evolution strategy that prioritizes correctness while optimizing performance. Finally, through extensive experiments, we demonstrate the effectiveness of CUDA-Harness, with further evaluations illustrating generalization across LLMs, hardware platforms, and to C-to-CUDA transpilation.
此外，我们提出了“反馈自适应演化”（Feedback-Adaptive Evolution），这是一种在优化性能的同时优先考虑正确性的内核演化策略。最后，通过广泛的实验，我们证明了 CUDA-Harness 的有效性，进一步的评估展示了其在不同 LLM、硬件平台以及 C 转 CUDA 转译任务中的泛化能力。