cuda-kernel-techniques

What problem does it solve?

This Skill helps you quickly choose the correct CUDA kernel optimization technique by bottleneck and context, reducing wasted experiments and regression churn.

Core Features & Use Cases

• Bottleneck-Driven Technique Lookup: Quickly route to the right sub-topic (memory access, data placement, parallelism, compute, control flow, occupancy, numerical stability, and anti-patterns) based on profiler symptoms (e.g., DRAM stalls, scoreboard stalls, occupancy limits).
• Hardware-Context Guidance (B200/SM100 + older SMs): Provides hardware notes and thresholds where behavior differs, so you can adapt techniques correctly for Blackwell B200 (SM100) and compare against SM80/SM90 expectations.
• Regression-Safe Debugging: Use the catalog’s dedicated anti-patterns and “when it hurts” sections to avoid known dead-ends when a change regresses.

Quick Start

Use the cuda-kernel-techniques skill to identify the best optimization technique for my kernel when NCU shows DRAM > 70% with long scoreboard stalls, and explain when cp.async and TMA help or hurt on SM100.