🤝 Top Collaborators
🤓 Latest Submissions
Kernel Olympics — CUDA→ROCm Migration Copilot
Problem AMD MI300X GPUs offer better price-performance than NVIDIA GPUs, but many enterprises remain locked into CUDA because about 20% of custom kernels cannot be migrated with hipify. These kernels contain warp-sensitive logic, shfl operations, and CUDA-specific APIs that require weeks of manual engineering, making migration the biggest barrier to ROCm adoption. Solution Kernel Olympics is a four-LLM pipeline that automates CUDA-to-ROCm migration using DeepSeek v4, GLM-5.2, Kimi-K2.7, and Gemma. It classifies kernels, generates HIP code, validates correctness, verifies execution on AMD hardware, and produces migration reports. Hipify-clang converts about 96.7% of CUDA syntax, while our classifiers detect high-risk patterns such as warp and wavefront divergence. A custom strip_to_kernel_only preprocessing step removes host code, improving translation accuracy. Generated code passes through three validation stages. The lexical gate rejects non-code output, the structural gate checks syntax and symbol preservation, and the semantic gate runs 31 tests with under 1% false positives. Failed outputs automatically retry with targeted feedback. A trigram-based pattern cache reuses previous solutions, reducing repeated migrations from about 12-second LLM calls to 0.2 ms lookups. Migrated kernels are compiled with hipcc on AMD MI300X using ROCm 7.2, executed, and compared byte-for-byte against CUDA outputs. Gemma then generates a plain-English migration summary. Key Innovations * Four-LLM autonomous workflow. * strip_to_kernel_only preprocessing. * Three-stage validation pipeline. * Pattern memory with 60,000× faster lookups. * Deterministic debug tracing. Performance Typical runtime is 105 seconds per kernel at $0.02-$0.05. The system migrates over 30 kernels per hour, performs 0.2 ms cache lookups, and maintains under 1% validation false positives.
13 Jul 2026
