Skip to content

Aditor v0.4.5

Choose a tag to compare

@ShiQu1218 ShiQu1218 released this 06 Nov 16:40
68918f1

新增與改善

  • 分析引擎影像運算加速:縮放、HSV、白色遮罩與長核形態學運算改為位元層級等價的更快實作。非 16:9 來源的縮放(1080p 縮至 720p)由 146.6 毫秒降至 13.0 毫秒(11.3 倍);480×118 白色遮罩由 550 微秒降至 72 微秒(7.7 倍);長核形態學由 11.0 毫秒降至 4.9 毫秒(2.3 倍)。1080p 即時偵測每影格約快 3%(16.19 → 15.67 毫秒)。
  • 觀戰畫面與分析前置加速:
    • 觀戰影格不再執行會被丟棄的 Mixtape 偵測並預先準備樣板,觀戰影格偵測由 21.99 毫秒降至 5.26 毫秒(4.2 倍),一般影格約快 4.7%。
    • 平行分析的分段改為平衡「實際解碼長度」:226 秒影片最長解碼段由 88.5 秒降至 80.6 秒(−9%),整體分析中位數 22.18 → 21.66 秒。
    • 未帶遊戲標籤的影片(匯入檔)其遊戲與畫面比例取樣改在同一解碼視窗內完成:前置階段由 4.38 秒降至 3.64 秒(−17%),整體分析 14.86 → 14.17 秒(−4.7%)。取樣影格逐位元相同,分析結果與修改前一致。
  • 錄影器喚醒次數降低:音訊混音迴圈由每秒 271 次喚醒降至 98 次,WGC 管理迴圈由每秒 9.0 次降至 0.99 次,讓筆電在錄影時 CPU 更容易進入深層閒置。
  • 遊戲偵測更省電:遊戲未執行時改為針對各遊戲的篩選查詢(約 31 毫秒),不再每 3 秒列出全部行程(約 500 毫秒 CPU,約佔單一核心 16.7%);遊戲執行中以輕量存活檢查取代,並每 30 秒完整比對一次以防範 PID 重用。
  • 匯入與編輯器儲存加速:內容未變的資料庫更新不再重寫並同步整份資料庫,也不再廣播更新。500 筆資料庫中,500 次未變更狀態更新由 1,350.6 毫秒降至 281.2 毫秒,100 次未變更精華更新由 206.6 毫秒降至 9.6 毫秒。

修正

  • 錄影啟動不再執行編碼器探測:錄影啟動時不再對每張 GPU 的每個編碼器做試編碼與音訊裝置列舉。先前在負載高的 Intel Arc 機器上,該探測會於 30 秒期限內在不同階段逾時,導致錄影重試持續失敗。
  • 監看失敗後可自動恢復:資料夾監看器失敗後會重新建立;Studio 相關 Hook 補齊遺漏的相依項目,避免使用過時狀態。
  • 錯誤回報涵蓋已攔截的錄影服務失敗:已被攔截處理的錄影服務失敗現在也會回報,協助診斷。

已知限制

  • 效能數據來自本機基準測試(AMD Ryzen 7 7700 與負載不一的開發機),不代表所有硬體;分析整體加速幅度僅約 2% 至 5%。
  • 遊戲行程查詢改為快取後,若 PID 被重用且查詢持續成功,「遊戲結束」的判定最多可能延遲約 30 秒外加一次防抖間隔。
  • 尚未驗證:1440p/超寬螢幕錄影的分析準確度、21:9/4:3 匯入影片的比例偵測、60 秒硬體錄影量測與遊戲 FPS 影響。
  • Windows 執行檔未進行程式碼簽章。

更新方式

  • 0.2.2 至 0.4.4 安裝版可透過自動更新升級。
  • 0.2.0/0.2.1 安裝版請手動下載並執行 Aditor-Setup-0.4.5.exe。
  • 免安裝版請下載 Aditor-0.4.5-Portable.exe 手動更新。

實驗數據:docs/v0.4.5-perf-experiments.md


New and improved

  • Faster engine image operations: resize, HSV, white mask and long-kernel morphology have bit-identical faster implementations. Non-16:9 resize (1080p to 720p) drops from 146.6 ms to 13.0 ms (11.3x); a 480x118 white mask from 550 us to 72 us (7.7x); long-kernel morphology from 11.0 ms to 4.9 ms (2.3x). Live detection is about 3% faster per 1080p frame (16.19 to 15.67 ms).
  • Faster spectating and analysis prelude:
    • Spectating frames no longer run a Mixtape detection whose result was discarded, and templates are prepared once: spectating-frame detection drops from 21.99 ms to 5.26 ms (4.2x), regular frames about 4.7% faster.
    • Parallel analysis segments now balance decoded length: the longest decoded segment of a 226 s clip drops from 88.5 s to 80.6 s (-9%); median analysis time 22.18 to 21.66 s.
    • Untagged files (imports) detect game and aspect in one decoder window: prelude 4.38 s to 3.64 s (-17%), full analysis 14.86 to 14.17 s (-4.7%). Sampled frames are bit-identical and results match the previous build.
  • Fewer recorder wakeups: the audio mixer loop drops from 271 to 98 wakeups per second and the WGC management loop from 9.0 to 0.99, letting laptop CPUs idle deeper while recording.
  • Cheaper game detection: with no game running, a filtered per-game query (~31 ms) replaces a full process listing every 3 s (~500 ms CPU, ~16.7% of one core); while a game runs, a lightweight liveness check is used with a full lookup every 30 s to guard against PID reuse.
  • Faster imports and editor saves: library updates that change nothing no longer rewrite, fsync or broadcast the whole library. In a 500-entry library, 500 unchanged status updates drop from 1,350.6 ms to 281.2 ms and 100 unchanged highlight updates from 206.6 ms to 9.6 ms.

Fixes

  • No encoder probe at recording start: starting a recording no longer trial-encodes every encoder on every GPU or enumerates audio devices. On a loaded Intel Arc machine that probe hit the 30 s deadline at a different stage each time, so recording retries kept failing.
  • Watcher recovery: failed folder watchers are recreated, and Studio hooks now declare their missing dependencies to avoid stale state.
  • Telemetry covers caught recording failures: recording service failures that were caught and handled are now reported for diagnosis.

Known limitations

  • Figures come from local benchmarks (AMD Ryzen 7 7700 and shared dev machines) and do not represent every configuration; end-to-end analysis gains are only about 2% to 5%.
  • With the cached process lookup, if a PID is reused while queries keep succeeding, the game-closed transition can be delayed by up to ~30 s plus one debounce interval.
  • Not yet verified: analysis accuracy on 1440p/ultrawide recordings, aspect detection on 21:9/4:3 imports, a 60 s hardware recording measurement, and game FPS impact.
  • Windows executables are not code signed.

How to update

  • 0.2.2 through 0.4.4 installs can upgrade through auto-update.
  • 0.2.0/0.2.1 installs need to download and run Aditor-Setup-0.4.5.exe manually.
  • Portable builds: download Aditor-0.4.5-Portable.exe to update manually.

Experiment data: docs/v0.4.5-perf-experiments.md

完整變更 / Full changelog: v0.4.4...v0.4.5