What problem does it solve? AscendC kernel developers on Ascend NPU hardware often pass correctness tests but miss performance targets, and manually interpreting msprof profiling CSVs to find bottlenecks is slow and error-prone. This Skill turns profiling data and operator code into concrete, evidence-backed optimization actions through a structured state-machine workflow. ## Core Features & Use Cases - Rule-first diagnosis with LLM fallback: Matches operator feature tags against an indexed library of expert optimization rules; when rule confidence is low, routes to deep pipeline analysis or algorithmic redesign subskills. - Closed-loop optimization workflow: Orchestrates INIT → TAG → SCORE → ROUTE → SUGGEST → APPLY → BUILD → EVALUATE → COMPARE → UPDATE phases with session isolation, resumable state, and automatic baseline comparison. - Knowledge capture: Verified optimizations are固化 back into the rule library via the rule_update subskill, so successful patterns become reusable. - Use Case: An operator like fused_adamw runs at 0.65x of the PyTorch baseline. Initialize the workspace from CAKE2 output, run the workflow, and receive ranked suggestions (tiling load balance, double buffering, UB fusion) that raise it to 2.48x with profiling-verified evidence. ## Quick Start Ask the assistant to run the code-performance-advisor workflow for your operator after initializing its workspace with init_workspace.py, for example: optimize the fastgelu operator using its latest profiling data.
