1478566 - Reduce doc.render() calls in webrender during video playback.

Status: RESOLVED DUPLICATE of bug 1474583

(Core :: Graphics: WebRender, defect, P3)

Description

[Sotaro Ikeda [:sotaro]]

Webrender calls doc.render() when display list is sent and when generate frame is requested. It seems redundant.

Comment 1

[Sotaro Ikeda [:sotaro]]

Attachment: patch - Reduce doc.render() calls in webrender

Comment 2

[Sotaro Ikeda [:sotaro]]

Attachment: patch - Reduce doc.render() calls in webrender

Comment 3

[Sotaro Ikeda [:sotaro]]

Attachment: patch - Reduce doc.render() calls in webrender

Comment 4

[Sotaro Ikeda [:sotaro]]

Attachment: patch another way - Reduce doc.render() calls in webrender

Comment 5

[Sotaro Ikeda [:sotaro]]

Attachment: patch another way - Reduce doc.render() calls in webrender

Comment 6

[Sotaro Ikeda [:sotaro]]

Attachment: patch another way - Reduce doc.render() calls in webrender

Comment 7

[Sotaro Ikeda [:sotaro]]

Attachment: patch - Reduce doc.render() calls in webrender

Comment 8

[Kartikaya Gupta (email:[e-mail])]

(In reply to Sotaro Ikeda [:sotaro] from comment #0)
> Webrender calls doc.render() when display list is sent and when generate
> frame is requested. It seems redundant.

It's not totally redundant. We need the first doc.render() for hit-testing correctly. If we get rid of the render we still need to apply the sampled frame messages and rebuild the clip scroll tree and hit-tester. I think this bug basically a duplicate of https://github.com/servo/webrender/issues/2807

Comment 9

[Kartikaya Gupta (email:[e-mail])]

Also, for the record - while reducing doc.render() calls will help with CPU usage, it will actually be worse for talos perf. I realized this after investigating the talos regressions from my WIP for servo/webrender#2807 (which is now sadly bitrotted). The reason is that doing two renders makes the second render cheaper than the first. So with two renders the time flow looks something like this:

   Vsync                                        Vsync                                        Vsync
   |--------------------------------------------|--------------------------------------------|
     \recvDL/ \scene build/ \expensive render/   \cheap render/

and without it, it would look like this:

   Vsync                                        Vsync                                        Vsync
   |--------------------------------------------|--------------------------------------------|
     \recvDL/ \scene build/ \redo hittester/     \expensive render/

So the time to the *end* of the final render increases. And that's what a lot of the talos tests measure (specifically, they wait for the MozAfterPaint which gets dispatched after the end of the render).

Comment 10

[Sotaro Ikeda [:sotaro]]

(In reply to Kartikaya Gupta (email:[e-mail]) from comment #8)
> (In reply to Sotaro Ikeda [:sotaro] from comment #0)
> > Webrender calls doc.render() when display list is sent and when generate
> > frame is requested. It seems redundant.
> 
> It's not totally redundant. We need the first doc.render() for hit-testing
> correctly. If we get rid of the render we still need to apply the sampled
> frame messages and rebuild the clip scroll tree and hit-tester. I think this
> bug basically a duplicate of https://github.com/servo/webrender/issues/2807

Thanks for the advice! But original intent of the bug is reduce doc.render() during video playback. It is related to Bug 1474583. So, the problem scope seems different than https://github.com/servo/webrender/issues/2807.

Comment 11

[Sotaro Ikeda [:sotaro]]

Attachment: patch - Reduce doc.render() calls in webrender

Comment 12

[Sotaro Ikeda [:sotaro]]

attachment 8995891 [details] [diff] [review] also causes talos regression :(

https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=42256b771d53533df3a9edba277bd7e4d191163f&newProject=try&newRevision=9c22b12c277465d21bce94354bad31d4b3831016&framework=1

Comment 13

[Sotaro Ikeda [:sotaro]]

Attachment: patch - Reduce doc.render() calls in webrender

Added PublishPipelineInfo handling.

Comment 14

[Sotaro Ikeda [:sotaro]]

(In reply to Sotaro Ikeda [:sotaro] from comment #12)
> attachment 8995891 [details] [diff] [review] also causes talos regression :(
> 
> https://treeherder.mozilla.org/perf.html#/
> compare?originalProject=try&originalRevision=42256b771d53533df3a9edba277bd7e4
> d191163f&newProject=try&newRevision=9c22b12c277465d21bce94354bad31d4b3831016&
> framework=1

The big regression seems to be addressed. In total, talos scores were improved.

https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=d8be8e68c4a79bd69ce0a36bdce3547aa85985e4&newProject=try&newRevision=8ec437e88eba2082ce8c9e49eee5e5e4d3a57c2e&framework=1

Comment 15

[Sotaro Ikeda [:sotaro]]

Attachment: patch - Reduce doc.render() calls in webrender

Comment 16

[Sotaro Ikeda [:sotaro]]

https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=00e262c79f6716ea4815ca83749da1b17d8e56de&newProject=try&newRevision=7cb9aff96b7b2a573e8f6724d86737344200adb4&framework=1

Comment 17

[Sotaro Ikeda [:sotaro]]

https://github.com/servo/webrender/pull/2994

Comment 18

[Sotaro Ikeda [:sotaro]]

This is going to be handled by Bug 1474583.