HIERA: Workload-Aware Planning Across Implementation Spaces for GPU Kernel Optimization

TL;DR - HIERA is a hierarchical planning framework that helps LLMs optimize GPU kernels by choosing among PyTorch operators, CUDA libraries, and custom CUDA implementations. It improves validity, sample efficiency, and optimization performance without additional model training.

• Uses contract-augmented task specifications to select an implementation space suited to each workload.
• Combines profiling feedback and expert knowledge for structured iterative refinement.
• Outperforms existing training-free methods on KernelBench across workload levels and base LLMs, while remaining competitive with training-based CUDA-L1.
• Achieves a 1.53× speedup over cuDNN on a specialized scientific-computing stencil operator.