What problem does it solve? PyPTO operators that pass correctness checks often still run far below hardware capability, and manual performance tuning on Ascend NPU requires expertise in swimlane analysis, TileShape configuration, graph merging, and scheduling. This Skill provides a structured, orchestrator-driven workflow that turns ad-hoc trial-and-error tuning into a controlled, measurable process with enforced accuracy gates. ## Core Features & Use Cases - Orchestrated tuning state machine: A tune-orchestrator enforces the fixed pipeline of environment checks, precision validation, performance data collection, analysis, phased tuning (frontend, swimlane, incore), and final report generation. - Phased optimization guidance: Dedicated sub-skills cover out-of-box code-level tuning (loop structure, TileShape, reshape), deep swimlane-based tuning (core usage, load balancing, graph merging, Stitch), and in-core tuning (L2 cache policy, instruction-level fixes, special shapes). - Automated performance analysis: The perf-analyzer component extracts core utilization, bubble rate, and load balance metrics from bubble_analysis.log and swimlane JSON files, then produces a rated performance report. - Use Case: Given a verified Add operator running at 27000 us with a 5x speedup target, the Skill validates precision on NPU, collects swimlane data, iterates single-parameter optimizations with per-round accuracy re-verification, and delivers a final tuning report with adopted and failed optimizations. ## Quick Start Ask the assistant to analyze and tune the performance of your implemented PyPTO operator, stating the operator directory and a concrete target such as a speedup multiple or maximum execution time.
