Presentation feedback on Windows uses GPUTimer which wraps command buffer queries and ultimately d3d11 fences through ANGLE.  This is suboptimal because it uses quite a bit of CPU (measured ~3% using ETW profiling).  It is also incorrect because completion of GPU work is not the same as presentation, and there's no way to determine latency after GPU work is complete.

I experimented with using DXGI_FRAME_STATISTICS from IDXGISwapChain::GetFrameStatistics(), and it looks like we can use it this way:
1) Num. of outstanding Presents = Num. of times Present() is called - stats.PresentCount (Num. of Presents() displayed on screen)
2) Present latency in DWM = stats.SyncRefreshCount - stats.PresentRefreshCount

There's some complexity because with DirectComposition the root surface switches between swap chain and IDCompositionSurface, and we can use get frame statistics only from swap chains.  Here's a prototype with support for EGLTimestampClient in DirectCompositionSurfaceWin using frame statistics with fallback when swap chain isn't used: https://chromium-review.googlesource.com/c/chromium/src/+/1266696

Making this work for non-DirectComposition path will involve more work, maybe exposing the swap chain via ANGLE?  geofflang@ jmadill@ thoughts?