[Webkit-unassigned] [Bug 252876] New: Slow performance of GPUP WebGL for rendering-heavy applications (like machine learning inference)

bugzilla-daemon at webkit.org bugzilla-daemon at webkit.org
Thu Feb 23 17:31:13 PST 2023


https://bugs.webkit.org/show_bug.cgi?id=252876

            Bug ID: 252876
           Summary: Slow performance of GPUP WebGL for rendering-heavy
                    applications (like machine learning inference)
           Product: WebKit
           Version: Safari 16
          Hardware: iPhone / iPad
                OS: iOS 16
            Status: NEW
          Severity: Normal
          Priority: P2
         Component: WebGL
          Assignee: webkit-unassigned at lists.webkit.org
          Reporter: tmullen.webkit at gmail.com
                CC: dino at apple.com, kbr at google.com, kkinnunen at apple.com

Observed so far on iPhone 12, 13 and 14, with iOS 16.2 and iOS 16.3.

Repro steps:
- Go to https://codepen.io/catchinglight/pen/JjaKvNQ
- Acknowledge warning and allow access to camera
- Point phone camera so as to keep exactly one hand visible in the selfie camera's field of view
- Observe FPS number
- Close this Safari tab
- Disable Settings app --> Safari --> Advanced --> Experimental Features --> Gpu Process: WebGL
- Open a new Safari tab
- Go back to https://codepen.io/catchinglight/pen/JjaKvNQ and observe much higher FPS number when repeating the same steps as before

The problem is not unique to the above CodePen, but the heavy use of WebGL there (for machine learning inference) coupled with an FPS display, makes it a nicer repro case than many other sites.

This bug seems possibly related: https://bugs.webkit.org/show_bug.cgi?id=250899, since the ML applications where this has been observed do use a readPixels call to grab GPU data back to CPU every frame for post-processing.  However, I'm not sure how to verify definitively one way or the other.

-- 
You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-unassigned/attachments/20230224/1f5b28d4/attachment.htm>


More information about the webkit-unassigned mailing list