cuda-roofline-strategy

What problem does it solve?

This Skill solves the problem of wasting GPU-kernel optimization experiments by choosing the next technique class without a principled diagnosis of what is currently limiting performance.

Core Features & Use Cases

• Classifies roofline bottlenecks (compute-bound, bandwidth-bound, occupancy-limited, latency-bound, balanced) from NCU metrics like SM throughput, DRAM/SOL throughput, achieved occupancy, and top warp stall reason.
• Selects a technique tier using iteration phase (early, mid, late, plateau) so recommendations scale with “how desperate” the run is and whether heavy rewrites are warranted.
• Supports per-workload refinement to avoid misleading averages by re-profiling slow outlier workloads and adapting dispatch when regimes differ.

Quick Start

Use this Skill when you have a fresh NCU profile and need to decide which category of CUDA kernel optimization to try next based on the current roofline position and your iteration phase.