1474985 - css blur is slow on this cnn.com slideshow

Status: RESOLVED DUPLICATE of bug 1434978

(Core :: Graphics, defect, P2)

Description

[Asa Dotzler [:asa]]

A report came in from twitter that CSS blur was prohibitively slow in Firefox. The testcase provided was this slideshow at cnn.com: https://www.cnn.com/ampstories/world/all-the-boys-are-out-photos-from-the-thai-cave-rescue

I profiled clicking through a few slides on the latest Windows nightly and here are the results: https://perf-html.io/from-addon/calltree/?hiddenThreads=6&thread=2&threadOrder=0-2-3-5-6-1-4&v=3

Comment 1

[Daniel Holbert [:dholbert]]

(In reply to Asa Dotzler [:asa] from comment #0)
> I profiled clicking through a few slides on the latest Windows nightly and
> here are the results:
> https://perf-html.io/from-addon/calltree/
> ?hiddenThreads=6&thread=2&threadOrder=0-2-3-5-6-1-4&v=3

That profiler URL is for a local (not-uploaded) profile -- any chance you've still got it handy (or could recapture) & could click the "share" link to get a sharable version?

Comment 2

[Maire Reavy [:mreavy]]

P2 to understand what's going on here.  Plus, it'd be nice to fix this one sooner than later.

Comment 3

[Asa Dotzler [:asa]]

Oops. Sorry for the bad profile link. Here's one that should work. https://perfht.ml/2NI2ilm

Comment 4

[Asa Dotzler [:asa]]

In the above profile, I advanced through 4 slides. You can see the red at each slide advance.

Comment 5

[Daniel Holbert [:dholbert]]

Thanks. For convenience, here's Asa's profile with just the cnn.com thread displayed: https://perfht.ml/2L5Ir15

There are definitely some long reflows shown there (50-100ms or longer). Looks like most of them involve grid reflow (nested several layers deep in some cases).  Here's that same profile, filtered for nsGridContainerFrame: https://perfht.ml/2L40Jjf

It looks like there are at least 3 nested grid containers here (3 nested calls to nsGridContainerFrame::Reflow).

Comment 6

[Daniel Holbert [:dholbert]]

I captured a profile on Linux, and I'm seeing very different results from Asa... Definite jank, but very little time spent in reflow, and quite a lot of time spent in the compositor (many back-to-back ~10-16ms composite operations) and in PaintThread.  Profile coming up.

Comment 7

[Daniel Holbert [:dholbert]]

Here's my profile on Linux: https://perfht.ml/2JcU8OE

I captured "PaintThread" as well (not present in profiles by default, hence not shown in Asa's profile).

As you can see, the jank bars correspond to time when the paint thread is active.  And most of that time seems to be spent in mozilla::gfx::AlphaBoxBlur::BoxBlur_SSE2

(I'm going to spin off a separate bug for the grid reflow jank that I initially noticed in Asa's profile, since I think that's unrelated to the main pain points reported on twitter about slow blurring in this example)

Comment 8

[Daniel Holbert [:dholbert]]

(I spun off bug 1475020 for the grid reflow issues.)

Comment 9

[Daniel Holbert [:dholbert]]

Seems like the BoxBlur stuff here is related to bug 1434978.

Comment 10

[Jamie Nicol [:jnicol]]

For me it runs really nicely with Direct2D, but yup it's painfully slow with Skia. It totally sucks in Chrome too. (But is great with webrender!) Do we know what platform the twitter user was on? Should they be getting D2D?

Everything else in the profile is completely eclipsed by the blur. And I don't know if we can speed that up. Bas?

Comment 11

[Asa Dotzler [:asa]]

(In reply to Jamie Nicol [:jnicol] from comment #10)
> For me it runs really nicely with Direct2D, but yup it's painfully slow with
> Skia. It totally sucks in Chrome too. (But is great with webrender!) Do we
> know what platform the twitter user was on? Should they be getting D2D?

Just asked and the reporter was on macOS.

Comment 12

[Daniel Holbert [:dholbert]]

(In reply to Jamie Nicol [:jnicol] from comment #10)
> It totally sucks in Chrome too. (But is great with webrender!)

--> marking as depending on bug 1475349 (metabug for performance fixed-by-webrender bugs)

Comment 13

[Daniel Holbert [:dholbert]]

(er sorry, bug 1464426 is the correct fixed-by-webrender metabug. marking as depending on that.)

Comment 14

[Bas Schouten (:bas.schouten)]

Since the reporter was on OS X, no, they should not be getting D2D :-).

Comment 15

[addieachan]

Hi! So I noticed this issue on Firefox Quantum on OS X, but not Firefox ESR on my Linux machine. I just wanted to confirm that at this point this is expected behavior, and if this is due to Skia as mentioned earlier. Also is there an ETA for a fix? Apologies if I was incorrect or unclear in anything. Thank you!