1677927 - (sw-wr-perf-blend-yuv) Very slow video decode for TikTok embeds (sw-wr)

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, defect)

Description

[Asif Youssuff]

Load https://www.vox.com/the-goods/2020/11/17/21570160/tiktok-selfie-pose-addison-charli and play each video one after another.

What happens: video very quickly goes out of sync and basically never recovers. Audio is fine.

Profile: https://share.firefox.dev/2IMaZNA

Comment 1

[Asif Youssuff]

Attachment: about:support

Comment 2

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

:yoasif, if you think that's a regression, then could you try to find a regression range in using for example mozregression?

Comment 3

[Jeff Muizelaar [:jrmuizel]]

Yeah, this should be improved by 1675990

Comment 4

[Jim Mathies [:jimm]]

Asif, are you still seeing this in latest Nightly?

Comment 5

[Asif Youssuff]

Still bad in the latest nightly - see new profile: https://share.firefox.dev/35FMLgU

Comment 6

[Jeff Muizelaar [:jrmuizel]]

Yeah, https://hg.mozilla.org/mozilla-central/rev/5e315caacc27 hasn't made it out into a nightly yet.

Comment 7

[Darkspirit]

Has this been fixed by bug 1675990 and bug 1678954?

Comment 8

[Asif Youssuff]

It is still not great.

Comment 9

[Lee Salzman [:lsalzman]]

Attachment: Bug 1677927 - avoid some unnecessary anti-aliasing work in YUV shader SWGL fast-path. r?jrmuizel

Comment 10

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/4d812a500d97
avoid some unnecessary anti-aliasing work in YUV shader SWGL fast-path. r=jrmuizel

Comment 11

[Natalia Csoregi [:nataliaCs]]

https://hg.mozilla.org/mozilla-central/rev/4d812a500d97

Comment 12

[Jim Blandy :jimb]

This is still pretty bad for me (e5e734f4fc19, 2021-1-21, including 4d812a500d97, but not 8e682a665334 for bug 1676554).

I'm seeing:

• 20% cs_clip_rectangle
• 30% (combined) brush_yuv_image_ALPHA_PASS_TEXTURE_RECT_YUV_frag
• 12% cs_clip_rectangle_FAST_PATH_frag

Comment 13

[Lee Salzman [:lsalzman]]

This fails to hit the fast-path in bug 1674618 because blending is used. linear_row_yuv could probably be adapted to support blending, but at the same time, performance here may be adequate as is so long as I can do some time to reduce the work we spend in cs_clip_rectangle.

Comment 14

[Lee Salzman [:lsalzman]]

Attachment: Bug 1677927 - Accelerate blended YUV in SWGL. r?jrmuizel

Despite having to move around some code into other files, this patch actually
changes very little inside composite.h and swgl_ext.h. It makes linear_row_yuv
able to call blend_pixels, and in turn makes blendYUV call that version of it.
This prevents cases where we have to blend video (because of clip masks) from
falling off the fast path.

To allow this, I had to move out blend stage code into blend.h so that it can
be called from inside composite.h. While I was at it, it made sense to factor
out the rasterization stage into its own rasterize.h file as well, as gl.cc
has grown significantly...

Comment 15

[Lee Salzman [:lsalzman]]

It seems like allowing acceleration of blended YUV leaves mostly the significant overhead of rendering the clip mask itself, which will probably speed up a bunch of other cases for us.

Comment 16

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/61b57740f175
Accelerate blended YUV in SWGL. r=jrmuizel

Comment 17

[Natalia Csoregi [:nataliaCs]]

https://hg.mozilla.org/mozilla-central/rev/61b57740f175

Comment 18

[Lee Salzman [:lsalzman]]

Comment on attachment 9205066 [details]
Bug 1677927 - Accelerate blended YUV in SWGL. r?jrmuizel

Beta/Release Uplift Approval Request

• User impact if declined: Trying to do partial SW-WR rollout on Linux. Would like to avoid people reporting significant performance regressions months down the line. This avoids embedded video being slow on some major sites like TikTok/Twitter/Facebook/etc.
• Is this code covered by automated tests?: Yes
• Has the fix been verified in Nightly?: Yes
• Needs manual test from QE?: No
• If yes, steps to reproduce:
• List of other uplifts needed: None
• Risk to taking this patch: Low
• Why is the change risky/not risky? (and alternatives if risky): Only impacts SW-WR/Linux. We can abort the rollout if unstable. Adequate time to deal with any bugs.
• String changes made/needed:

Comment 19

[Ryan VanderMeulen [:RyanVM]]

Comment on attachment 9205066 [details]
Bug 1677927 - Accelerate blended YUV in SWGL. r?jrmuizel

Approved for 87.0b4.

Comment 20

[Ryan VanderMeulen [:RyanVM]]

Comment on attachment 9205066 [details]
Bug 1677927 - Accelerate blended YUV in SWGL. r?jrmuizel

https://hg.mozilla.org/releases/mozilla-beta/rev/8a7a300d0c6f

Removing the approval to get this off the needs-uplift radar since this bug is marked leave-open.

Comment 21

[Lee Salzman [:lsalzman]]

With bug 1682194, the remaining clip rectangle overhead is mostly resolved. Closing this out to make room for new perf bugs.