What problem does it solve? AscendC operator development produces scattered artifacts—precision results, msprof profiling CSVs, kernel source, and tiling constants—that are hard to review together. This Skill consolidates them into one self-contained HTML dashboard with four tabs: algorithm flow, UB memory/tiling, precision analysis, and performance report. ## Core Features & Use Cases - Two-stage generation: gen_dashboard.py extracts objective data (precision results, profiling CSVs, UB buffers, tiling constants) into panels/*/data.json, then the AI writes visualization fragments (flow.html, steps.html, ub_viz.html) and analysis text before reassembling the final dashboard. - Quality validation: check_dashboard.py runs structural, data, coverage, value, contract, rendering, and health checks on the generated HTML, reporting PASS/WARN/FAIL per item. - Works with incomplete results: dashboards can be generated as soon as compilation passes, displaying FAIL case diagnostics and performance warnings when precision has not passed. - Use Case: After implementing a BatchMatmulMaxSum kernel on Ascend NPU, point the Skill at output/<op_name>/ to produce an offline-shareable dashboard showing the Cube/Vector compute flow, UB allocation donut chart, per-case precision errors, and speedup charts. ## Quick Start Generate an operator dashboard from my AscendC output directory output/MyOp/ and validate it with the quality checker.
