1569881 - Investigate a workaroud for Flickering selection/text input(Win10 (1903)/Intel HD Graphics 530 + 620)

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, task)

Description

[Sotaro Ikeda [:sotaro]]

Chromium uses IDXGISwapChain1::Present1 () for presenting a frame. It is better to use it also in RenderCompositorANGLE.

https://cs.chromium.org/chromium/src/ui/gl/direct_composition_child_surface_win.cc?l=166

Comment 1

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch - Use IDXGISwapChain1::Present1() in RenderCompositorANGLE

Comment 2

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch - Use IDXGISwapChain1::Present1() and allow tearing in RenderCompositorANGLE

Comment 3

[Sotaro Ikeda [:sotaro]]

Attachment 9081551 [details] [diff] did not address the problem of Bug 1556634. But Attachment 9081552 [details] [diff] address the problem of Bug 1556634 for me.

Comment 4

[Sotaro Ikeda [:sotaro]]

:jrmuizel, can you check if Attachment 9081552 [details] [diff] works for you?

Comment 5

[Jeff Muizelaar [:jrmuizel]]

(In reply to Sotaro Ikeda [:sotaro] from comment #4)

:jrmuizel, can you check if Attachment 9081552 [details] [diff] works for you?

It does work for me.

However, from my reading of the Chrome code it looks like DXGI_SWAP_CHAIN_FLAG_ALLOW_TEARING is not used by default. This suggests that there must be some other reason that Chrome works and Firefox doesn't.

Comment 6

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch - Use DXGI_ALPHA_MODE_PREMULTIPLIED

DXGI_ALPHA_MODE_PREMULTIPLIED usage also addressed the problem for me.

Comment 7

[Sotaro Ikeda [:sotaro]]

Attachment 9081847 [details] [diff] and Attachment 9081552 [details] [diff] addressed the problem for me. But performance seemed to become worse.
FPS of the following WebGL was degraded.

• https://akirodic.com/p/jellyfish/

Comment 8

[Sotaro Ikeda [:sotaro]]

:jrmuizel, can you check if Attachment 9081847 [details] [diff] also works for you?

Comment 9

[Jeff Muizelaar [:jrmuizel]]

Attachment 9081847 [details] [diff] also works for me.

Comment 10

[Sotaro Ikeda [:sotaro]]

Attachment: temporal patch - Update only part of back buffer if possible

When partial present did not render entire back buffer, the problem seemed to go away, though part of the window was not rendered.

Comment 11

[Sotaro Ikeda [:sotaro]]

Attachment: Bug 1569881 - Short term workaround for Intel GPU

Comment 12

[Sotaro Ikeda [:sotaro]]

:jrmuizel, do we need to land PREMULTIPLIED workaround Attachment 9084980 [details] even when it regresses MotionMark performance(180 -> 70)?

Comment 13

[Jeff Muizelaar [:jrmuizel]]

I'm still inclined to take the workaround. I think we might need something better to actually ship on Intel but I think this is good enough for Nightly and maybe Beta

Comment 14

[Sotaro Ikeda [:sotaro]]

Ok, thank you :)

Comment 15

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5d16fc1cebc7
Short term workaround for Intel GPU r=jrmuizel

Comment 16

[Sebastian Hengst [:aryx] (needinfo me if it's about an intermittent or backout)]

https://hg.mozilla.org/mozilla-central/rev/5d16fc1cebc7

Comment 17

[Jeff Muizelaar [:jrmuizel]]

Alexis had the idea of using partial present with 2 rects that cover the whole screen instead of one rect. If that doesn't work we also have the option of using 2 rects and leaving one pixel unexposed.

Comment 18

[Sotaro Ikeda [:sotaro]]

(In reply to Jeff Muizelaar [:jrmuizel] from comment #17)

If that doesn't work we also have the option of using 2 rects and leaving one pixel unexposed.

What could we get with it? Forced partial present with current WebRender also regressed performance.

Comment 19

[Jeff Muizelaar [:jrmuizel]]

Does it regress performance the same amount as using ALPHA?

Comment 20

[Sotaro Ikeda [:sotaro]]

I am going to compare them.

Comment 21

[Sotaro Ikeda [:sotaro]]

Attachment: temporal patch - Update only part of back buffer if possible

Comment 22

[Sotaro Ikeda [:sotaro]]

Attachment: temporal patch - Update only part of back buffer if possible

Comment 23

[Sotaro Ikeda [:sotaro]]

Attachment: temporal patch - Use partial present with 2 rects

Correct patch.

Comment 24

[Sotaro Ikeda [:sotaro]]

(In reply to Jeff Muizelaar [:jrmuizel] from comment #17)

Alexis had the idea of using partial present with 2 rects that cover the whole screen instead of one rect.

Attachment 9085653 [details] [diff] implemented it. But it did not address the problem.

Comment 25

[Sotaro Ikeda [:sotaro]]

Attachment: temporal patch - Use partial present and one pixel unexposed

Comment 26

[Sotaro Ikeda [:sotaro]]

(In reply to Jeff Muizelaar [:jrmuizel] from comment #17)

If that doesn't work we also have the option of using 2 rects and leaving one pixel unexposed.

Attachment 9085653 [details] [diff] implemented it. It addressed the problem for me.

Comment 27

[Sotaro Ikeda [:sotaro]]

I compared MotionMark score of the followings on my laptop. [1] was best performance, then [2] was better than [3]

[1] No workaround : 160-190
[2] Using 2 rects and leaving one pixel unexposed : 100-130
[3] Use DXGI_ALPHA_MODE_PREMULTIPLIED : 80-90

---

build of [1]

• https://treeherder.mozilla.org/#/jobs?repo=try&revision=f6bd14f93600774fb65faba604cf1e9769c5b170
• https://treeherder.mozilla.org/#/jobs?repo=try&revision=08dae7f6ddc7338b7621e91da2d0292fbded1e9c

build of [2]

• https://treeherder.mozilla.org/#/jobs?repo=try&revision=012311ac5f3496e1c21d232dfa77d171985b0807
• https://treeherder.mozilla.org/#/jobs?repo=try&revision=3c8ecb14a7b4322d81f07bfc3591ea16877b844f

build of [3]

• https://treeherder.mozilla.org/#/jobs?repo=try&revision=a99450cc87300859fc0215861f89e1a96b3245ab
• https://treeherder.mozilla.org/#/jobs?repo=try&revision=865081864bc0cc9a0c0db5b91987e3524092dd82