1451112 - Scrolling of specific website very slow with WebRender enabled (http://www.pcgameshardware.de)

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, defect, P2)

Description

[Clemens Eisserer]

User Agent: Mozilla/5.0 (X11; Linux x86_64; rv:61.0) Gecko/20100101 Firefox/61.0
Build ID: 20180403100105

Steps to reproduce:

1. Loaded:  http://www.pcgameshardware.de/Notebook_Laptop-Hardware-201330/Specials/Ryzen-Mobile-im-Vergleich-1252536/#a3
2. Scrolled down a bit


Actual results:

When reaching the "blurred" content (paywall), scrolling starts to become choppy with WebRender enabled - while with the default OpenGL compositor it works well.

However, I wasn't really able to spot the issue in the profile: https://perfht.ml/2GR7XVS
The firefox-bin process consumes 100% cpu, while the GPU profiler shows the GPU is basically idle.

Comment 1

[Clemens Eisserer]

some more details about the system: AMD Kaveri 7650k APU, Mesa-17.3.6, Linux-4.15.14

Comment 2

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

From the profile it looks like the blob image rendering is blocking the render backend thread, which then blocks APZ hit-testing, which in turn blocks the compositor thread. APZ assumes the RB thread is going to be responsive, so from my point of view the blocking on expensive blob image operations is bad.

Comment 3

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

This might actually become a serious problem - Jeff, do we have plans to make blob image rendering not block the render backend thread? If not, and if we can't guarantee a responsive RB thread, then maybe we should go back to using gecko hit-testing rather than using WR hit-testing. Or find a way to do WR hit-testing without blocking on the RB thread.

Comment 4

[Jeff Muizelaar [:jrmuizel]]

We do have plans to make blob image rendering not block the render backend. See the just filed bug 1455422.

Comment 5

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

Thanks

Comment 6

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

Also for the record, I can repro this, and it happens even with async scene building enabled.

Comment 7

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

For me this is "better" now in that we don't jank so much, but checkerboard instead. Will leave this open for now since we can probably still improve the perf here.

https://perfht.ml/2O7SzFb

Comment 8

[Jeff Muizelaar [:jrmuizel]]

Runs well enough on Nvidia hardware

Comment 9

[Clemens Eisserer]

Indeed a lot better compared to when I reported it.
Runs well here on an AMD APU on Linux using the RadeonSI driver, too.