| Summary: | WinCairo scores 1.00 on MotionMark benchmark since r261113 | ||
|---|---|---|---|
| Product: | WebKit | Reporter: | Fujii Hironori <Hironori.Fujii> |
| Component: | Platform | Assignee: | Nobody <webkit-unassigned> |
| Status: | NEW --- | ||
| Severity: | Normal | CC: | sabouhallawa |
| Priority: | P2 | ||
| Version: | WebKit Nightly Build | ||
| Hardware: | Unspecified | ||
| OS: | Unspecified | ||
| URL: | https://browserbench.org/MotionMark1.1/ | ||
| See Also: |
https://bugs.webkit.org/show_bug.cgi?id=211431 https://bugs.webkit.org/show_bug.cgi?id=211582 |
||
| Attachments: | |||
|
Description
Fujii Hironori
2020-05-10 21:29:49 PDT
Created attachment 398997 [details]
Patch not to use WM_TIMER
WM_TIMER seems the bottleneck.
Is the cause of the MotionMark regression r261113 or r261264 and r261333? I think r261113 was affecting the iOS Low power case only. But after r261264, scheduling the RenderingUpdate on Windows goes through DisplayRefreshMonitor which is different from before. I have a WinCairo bulit binary of r261262. It already has the same issue. But, I'm not sure which revision is the real culprit. Created attachment 399200 [details] MotionMark of WinCairo WK1 (r261577) r261577 fixed another rAF issue of WinCairo WK1, I tried MotionMark with WinCairo WK1. WinCairo WK1 works well for some tests, but scores 1.00 only for Focus and Images tests. Those tests are run in Accelerated Compositing mode. WinCairo WK1 AC is using a timer. https://trac.webkit.org/browser/webkit/trunk/Source/WebKitLegacy/win/WebCoreSupport/AcceleratedCompositingContext.cpp?rev=261577#L274 AppleWin WK1 AC is using PostThreadMessageW. https://trac.webkit.org/browser/webkit/trunk/Source/WebCore/platform/graphics/ca/win/LayerChangesFlusher.cpp?rev=261577#L119 Created attachment 399227 [details] MotionMark of WinCairo WK1 (r261577) (No updateRendering in WM_PAINT) Why does WinCairo WK1 non-AC perform well for MotionMark even though it's also using a timer? The reason seems that it calls Page::updateRendering twice in a single timer fire. Once in WebView::paint, another in WebView::flushPendingGraphicsLayerChangesSoon. I tried MotionMark with WinCairo WK1 again by applying a patch not to call Page::updateRendering in WebView::paint. MotionMark score falled down to 1.39. (I don't know why it isn't 1.00) diff --git a/Source/WebKitLegacy/win/WebView.cpp b/Source/WebKitLegacy/win/WebView.cpp index fe7cb6a1d03..ec44027e06c 100644 --- a/Source/WebKitLegacy/win/WebView.cpp +++ b/Source/WebKitLegacy/win/WebView.cpp @@ -1285,7 +1285,7 @@ void WebView::paint(HDC dc, LPARAM options) { LOCAL_GDI_COUNTER(0, __FUNCTION__); - m_page->updateRendering(); + //m_page->updateRendering(); if (paintCompositedContentToHDC(dc)) { ::ValidateRect(m_viewWindow, nullptr); Created attachment 399240 [details]
html to console.log the interval of requestAnimationFrame
This html shows the duplicated frame of WinCairo WK1 non-AC.
Chromium seems to be using own message pump. Created attachment 399295 [details]
experimental patch to conditionally use a timer or callOnMainThread depending on the two-frame-previous tick
FPS can't be stable with this patch.
|