roofline

What problem does it solve? When profiling an LLM serving workload on AMD Instinct GPUs, the biggest kernel by GPU time is not always the one with the most optimization headroom. This Skill answers how close each top kernel already is to its hardware roofline ceiling, so optimization budget goes to kernels with real recoverable gains instead of saturated ones. ## Core Features & Use Cases - Roofline headroom analysis: Computes achieved bandwidth/FLOPs against hardware peaks for each top kernel, classifying it as memory-bound, compute-bound, or latency-bound, and emitting attainable speedup and expected end-to-end gain. - Byte/FLOP models per op class: Provides analytic models for dense GEMM, MoE grouped GEMM, paged attention decode, and elementwise kernels, with confidence stages that sharpen as real shapes and rocprofv3 counters become available. - Non-fatal degradation ladder: Every failure mode (missing peaks, unmodelable op class, infeasible ratio, unavailable counters) degrades gracefully so the profiling run never fails because of this analysis. - Use Case: Given a vLLM decode profile where a fused MoE kernel takes 26% of GPU time and paged attention takes 9%, the Skill shows the MoE kernel is saturated at 88% of the memory roofline while attention is underperforming with 1.7-2.8x headroom, reversing the naive priority order. ## Quick Start Analyze the top kernels in my latest GPU profile and tell me which ones have real optimization headroom versus which are already saturated at the roofline.