KernelPro Optimizes GPU Kernels with LLMs and Micro-Profiling

Optimizing GPU kernel code is a highly specialized and time-consuming task, typically requiring expert human intervention. This research introduces KernelPro, a novel closed-loop multi-agent system designed to automate and significantly enhance this process. KernelPro combines the code generation capabilities of Large Language Models (LLMs) with detailed hardware profiler feedback and specialized bottleneck detection tools. The system's innovative design includes a semantic feedback operator that translates raw hardware metrics into actionable natural language guidance, effectively acting as an expert surrogate. It employs a two-stage tool invocation architecture, where a roofline-based classifier intelligently selects which specialized profiling tools (like `ncu`, SASS, `nsys`) to deploy for kernel, instruction, or system-level analysis. Furthermore, KernelPro utilizes a domain-adapted Monte Carlo Tree Search (MCTS) algorithm with progressive widening and other enhancements to guide its code optimization process, including direct CuTe source-level code generation. KernelPro achieved state-of-the-art performance on KernelBench, demonstrating geometric mean speedups of up to 5.30x on challenging optimization levels. Notably, it even surpassed expert-optimized MoE training kernels from VeOmni by 1.23x, generating a raw-CUDA+CuTe Hopper WGMMA kernel from scratch. Beyond speed, KernelPro is also the first system of its kind to optimize for energy efficiency, achieving an 11.6% measured energy reduction at matched speed, highlighting the significant impact of its micro-profiling tools, MCTS search, and proactive tool orchestration.