Iteratively optimize a CUDA/CUTLASS/Triton kernel against a reference implementation using ncu-guided reasoning. Use this skill whenever the user asks to optimize, speed up, or improve the performance of a .cu kernel (CUDA or CUTLASS) or a Triton/Python kernel file, especially when they provide a baseline operator and a reference, mention "ncu", "Nsight Compute", "iterative optimization", "kernel tuning", or ask Claude to "make this kernel faster". The skill drives a multi-iteration roofline-guided optimization loop: profile with ncu → compute roofline gaps → allocate axis budgets → pick methods by priority scan → generate K branch candidates → validate + benchmark → select champion → ablation attribution → SASS verification → update global state. Each iteration's artifacts (kernel, CoT analysis, ncu-rep) are persisted under a timestamped run folder, and a final summary is emitted.
2026-04-22