msopprof-visualization

Collect msOpProf operator profiling data and generate standalone HTML performance reports.

Updated Sep 15, 2026
One-click install
npx skills add https://github.com/WangWindow/CANN-BatchMatMulMaxsum --skill msopprof-visualization-wangwindow
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: msopprof-visualization
Source: https://github.com/WangWindow/CANN-BatchMatMulMaxsum/tree/main/.agents/skills/ops-profiling/msopprof-visualization
Command: npx skills add https://github.com/WangWindow/CANN-BatchMatMulMaxsum --skill msopprof-visualization-wangwindow

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve? Profiling Ascend NPU operators with msOpProf requires many manual profiler replays, environment checks, and artifact inspections, and the raw outputs are hard to interpret. This Skill automates environment resolution, minimal non-duplicated collection, semantic artifact validation, and generation of a single interactive HTML report. ## Core Features & Use Cases - Minimal Collection Planning: Maps requested report pages (Details, Roofline, Timeline, Cache, Source, Warp Stall, Instruction Timeline, Raw Data, On-Chip Memory) to the fewest profiler replays, with semantic reuse and aliasing of completed collections. - Semantic Validation: Verifies that zero-exit profiler runs actually produced usable artifacts (CSV suites, trace events, source line maps, debug_line data) and records actionable failure reasons instead of fabricating metrics. - Standalone HTML Reports: Builds one canonical payload rendered into an interactive report with per-module availability diagnostics, plus machine-readable report_index.json and collection_manifest.json. - Use Case: After developing an Ascend C operator kernel, run the full pipeline against the operator project to obtain a complete performance report, or regenerate targeted pages (e.g. roofline and cache) from an existing collection directory without re-profiling. ## Quick Start Ask the assistant to profile the operator project at a given path and generate a complete msOpProf HTML performance report in a chosen output directory.

Frequently Asked Questions about msopprof-visualization

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I generate an msOpProf performance report for an Ascend operator?

Run the pipeline script with the operator project path and an output directory, choosing a preset such as complete. It resolves the CANN environment and executable, runs the minimal profiler replays, validates artifacts, and writes a standalone HTML report plus report_index.json.

How to rebuild a report from existing msprof collection results?

Use visualize mode pointed at a directory containing collection_manifest.json. The Skill reuses the recorded artifacts without re-running the profiler and regenerates the HTML report, optionally restricted to specific feature pages.

Why is the Source page unavailable in my profiling report?

Source mapping requires .debug_line sections in the exact profiled executable's build tree. Rebuild with line-table debug data (-g), confirm the rebuilt binary is the one being profiled, and rerun the Source collection.

Does the collector re-run profiler commands that already succeeded?

No. Completed blocks are fingerprinted by command, application, and collector version, and semantically complete artifacts are reused or aliased across modules. Results from timed-out, failed, or semantically empty runs are never reused.

What happens when a profiling module cannot be collected?

The report marks the module unavailable with an evidence-based reason and a suggested next action, rather than fabricating metrics. Other modules still render normally from the same canonical payload.