codeswarm

Map CUDA kernel optimization strategies to HIP/ROCm implementations with performance-focused rewrites. Use when a CUDA kernel has already been translated (or partially translated) and needs optimization parity on AMD GPUs, especially for tiling, pipelining, shared-memory/LDS staging, warp-to-wavefront adaptation, vectorized loads/stores, and tensor-op path selection.

Convert CUDA source to HIP-compatible source with minimal semantic drift. Use when a task involves `.cu`, `.cuh`, or CUDA-specific C++ code that must compile and run under ROCm/HIP, including API replacement, intrinsic translation, launch syntax conversion, and identifying CUDA-only features that require algorithmic rewrites.