1488980 - nsIDOMWindowUtils::startFrameTimeRecording isn't implemented for WebRender

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, enhancement, P1)

Description

[Matt Woodrow (:mattwoodrow)]

This is used for tscrollx and TART talos tests, and is silently recording 0's when WR is enabled.

Comment 1

[Bobby Holley (:bholley)]

(In reply to Matt Woodrow (:mattwoodrow) from comment #0)
> This is used for tscrollx and TART talos tests, and is silently recording
> 0's when WR is enabled.

Nice find! To clarify, is that making WR numbers better or worse than they should be?

Comment 2

[Matt Woodrow (:mattwoodrow)]

It makes the WR numbers look better than they are unforunately.

The main-thread scrolling results appear to be fairly comparable, with the exception of the svg subtest, which is almost an order of magnitude worse.

The APZ scrolling results are all 0s, so it roughly halves the combined average for the suite.

Comment 3

[Matt Woodrow (:mattwoodrow)]

Attachment: Bug 1488980 - Implement StartFrameTimeRecording for WebRender. r?mstange

MozReview-Commit-ID: Ib5t0e8S6df

Comment 4

[Matt Woodrow (:mattwoodrow)]

The attached patch fixes all the APZ subtests except for iframe.svg, which I filed bug 1488996 for.

Comment 5

[Bobby Holley (:bholley)]

(In reply to Matt Woodrow (:mattwoodrow) from comment #2)
> The main-thread scrolling results appear to be fairly comparable, with the
> exception of the svg subtest, which is almost an order of magnitude worse.

Huh, that seems worth looking into. My understanding is that svg blob rasterization is slow, but once that's done, it seems like we should be able to blit it pretty fast on scrolls. I wonder what's going on?

Comment 6

[Matt Woodrow (:mattwoodrow)]

(In reply to Bobby Holley (:bholley) from comment #5)
> (In reply to Matt Woodrow (:mattwoodrow) from comment #2)
> > The main-thread scrolling results appear to be fairly comparable, with the
> > exception of the svg subtest, which is almost an order of magnitude worse.
> 
> Huh, that seems worth looking into. My understanding is that svg blob
> rasterization is slow, but once that's done, it seems like we should be able
> to blit it pretty fast on scrolls. I wonder what's going on?

The scrolling thing is inside the blob, which is why it doesn't work with APZ currently. Bug 1449634 is going to fix that.

My guess is that the same issue is causing us to re-rasterize the blob on each scroll, which would be painfully slow, but also might be fixed by the same patch. Doing a local build now to confirm.

Comment 7

[Matt Woodrow (:mattwoodrow)]

Indeed, the patch from bug 1449634 gives us a score for APZ scrolling, and improves the main thread scrolling score significantly.

Comment 8

[Markus Stange [:mstange]]

Comment on attachment 9006753 [details]
Bug 1488980 - Implement StartFrameTimeRecording for WebRender. r?mstange

Markus Stange [:mstange] has approved the revision.

Comment 9

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/6d3db9b4da32
Implement StartFrameTimeRecording for WebRender. r=mstange

Comment 10

[Natalia Csoregi [:nataliaCs]]

https://hg.mozilla.org/mozilla-central/rev/6d3db9b4da32