1681747 - sw-wr performance bad when viewing pdfs on macOS

Status: REOPENED

(Core :: Graphics: WebRender, defect)

Description

[Jeff Muizelaar [:jrmuizel]]

https://share.firefox.dev/2W2r3yc

We're spending lots of time in composite_TEXTURE_RECT_frag::run which isn't fast-pathed.

Comment 1

[Lee Salzman [:lsalzman]]

Attachment: Bug 1681747 - add SWGL fast-path for composite shader. r?jrmuizel

Comment 2

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5f82b68a19d5
add SWGL fast-path for composite shader. r=jrmuizel

Comment 3

[Alexandru Michis [:malexandru]]

https://hg.mozilla.org/mozilla-central/rev/5f82b68a19d5

Comment 4

[Alexandru Ionescu (needinfo me) [:alexandrui]]

== Change summary for alert #28087 (as of Tue, 15 Dec 2020 12:29:47 GMT) ==

Improvements:

Ratio	Suite	Test	Platform	Options	Absolute values (old vs new)
18%	glterrain		macosx1014-64-shippable-qr	e10s stylo webrender-sw	10.64 -> 8.70
18%	glterrain		macosx1014-64-shippable-qr	e10s stylo webrender-sw	10.70 -> 8.75

For up to date results, see: https://treeherder.mozilla.org/perfherder/alerts?id=28087

Comment 5

[Jeff Muizelaar [:jrmuizel]]

This helped but performance is still not amazing: https://share.firefox.dev/2Kx4aQM

Comment 6

[Lee Salzman [:lsalzman]]

We should investigate why on macOS we're going to the composite shader in this case instead of actually going through the native compositor to composite this directly.

Comment 7

[Markus Stange [:mstange]]

The "WR OS compositor" markers on the Renderer thread in Jeff's profiles show that the native compositor clearly is used. So maybe we use the composite shader to composite external images like canvases into picture tiles?

Comment 8

[Glenn Watson [:gw]]

If an external image doesn't get attached as a native surface, it will fall back to using the composite shader to construct a native surface from whatever the external data source is. However, it's not clear to me why this would be happening on mac from a quick look at the code.

I'm currently looking into fixing https://bugzilla.mozilla.org/show_bug.cgi?id=1679480, and I think a good first step for that may be to refactor the Draw compositor to be implemented in terms of the normal compositor trait (this will greatly simplify the picture cache code for managing surfaces, which should make it much simpler to fix the referenced bug). When I get to that, it may become apparent why we're hitting the composite shader in this case, if no one else gets to investigate it first.

Comment 9

[BugBot [:suhaib / :marco/ :calixte]]

The bug assignee didn't login in Bugzilla in the last 7 months, so the assignee is being reset.