[Webkit-unassigned] [Bug 90375] Parallel image decoders

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=90375





--- Comment #17 from Kwang Yul Seo <skyul at company100.net>  2012-07-12 02:43:23 PST ---
(In reply to comment #16)
> (In reply to comment #13)
> 
> > As mentioned above, parallel image decoders can't improve page loading time (or first paint time) dramatically if images are not cached, because it takes a long time to fetch large images and image decoding can be performed in the meantime (interleaved with loading). If images are already cached, then parallel image decoders shine! 
> 
> The abstract of your design document suggests you consistently outperform sequential image decoding.  Yet small images are excluded from threaded decodes for some reason, so perhaps that means that most web pages would be excluded?  

Yes, small images are slow due to synchronization overhead. We are still experimenting with various threshold values. We've just found that the break even point is near 128x128. According to Zoltan's survey https://bugs.webkit.org/show_bug.cgi?id=71555#c7, it looks many sites can get performance benefits from parallel image decoders. Of course, the improvement is marginal near 128x128.

>Another result is that large image decodes are network bound, and also see no benefit.  I don't see how you can "consistently outperform" the sequential image decoders.  Perhaps the improvement is marginal at best?  

Sadly, yes. However, I see other use cases. What about Offline Web Applications or Widgets? Because images are locally stored, parallel image decoders can speed up initial loading time.

By "consistently outperform", I meant parallel image decoders are used only when they can do better based on various criteria. Otherwise, serial image decoders are used instead with no switching overhead.

>It state that the cached image case "shines" (I don't know what that means) but it is the only case that benefits in some way.  Why should I care about that case?

We do understand that parallelization usually introduces massive code changes with only marginal performance improvements. Our test results show that parallel image decoders are faster in some cases, and at least not slower in most cases.

We are trying to further improve performance and reduce the complexity. So I hope we could discuss this in a more positive direction.

-- 
Configure bugmail: https://bugs.webkit.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.