1414575 - PeerPad has a really bad time with WebRender

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, defect, P4)

Description

[Jeff Muizelaar [:jrmuizel]]

https://peerpad.net/#/

Comment 1

[Jeff Muizelaar [:jrmuizel]]

This is probably blur related. We should retry after https://github.com/servo/webrender/pull/1961 lands.

Comment 2

[Nicolas Silva [:nical]]

I believe #1961 landed and this page is still pretty bad (and renders incorrectly).

Comment 3

[Markus Stange [:mstange]]

This is indeed blur related. The two <a> elements for "IPFS" and for "CRDTs" each have an ::after pseudo element which has box-shadow: -350px -20px 1000px 22px #23e0f7 set on it. That's a 1000px radius blur.

Comment 4

[Ethan Lin[:ethlin]]

I take a look.

Comment 5

[Ethan Lin[:ethlin]]

Looks like the performance bottleneck is not box-shadow anymore. I sent a PR 2096[1] to fix the rendering problem. I will keep investigating the next performance bottleneck in this page.

[1] https://github.com/servo/webrender/pull/2096

Comment 6

[Ethan Lin[:ethlin]]

I should say the blur is not the bottleneck. According to my local test, we spend 25ms on GPU, and clip takes more than 10ms. I think it's because the clip shape is ellipse and webrender doesn't handle ellipse clip efficiently. I'll have a PR to fix it.

Comment 7

[Ethan Lin[:ethlin]]

https://github.com/servo/webrender/pull/2103

Comment 8

[Ethan Lin[:ethlin]]

PR 2158[1] also helps. Right now the GPU time is about 11ms on my MBP.

[1] https://github.com/servo/webrender/pull/2158

Comment 9

[Ethan Lin[:ethlin]]

The FPS of scrolling on my device is about 50~60fps now. Jeff, do you think we can close this bug?

Comment 10

[Jeff Muizelaar [:jrmuizel]]

I only get around 30fps on that page and see 30+ ms composites: https://perfht.ml/2nE69al

Comment 11

[Jeff Muizelaar [:jrmuizel]]

This is better but still not great.

Comment 12

[Maire Reavy [:mreavy]]

Jeff -- What issues remain?  How bad are they?  How difficult are they to hit?

Comment 13

[Jeff Muizelaar [:jrmuizel]]

I tried this out on some older nvidia desktop hardware and still got 30ms composites. So we could definitely do better.

Comment 14

[Jeff Muizelaar [:jrmuizel]]

This seems to work reasonably well now.