1575159 - Partial Invalidation on Windows

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, enhancement, P3)

Description

[Sotaro Ikeda [:sotaro]]

Once we have the infrastructure in WR for partial rects. We could implement partial invalidation on Windows. Partial invalidation is also necessary for android.

Comment 1

[Sotaro Ikeda [:sotaro]]

• It seems reasonable to use infrastructure of WR Picture caching to calculate invalidated rects for Partial Invalidation. But current picture caching does not cover whole window rendering. Bug 1575767 might be related.
• WR partial rendering should update only within invalidated rects.
• Invalidated rects might needs to be simple if we want to get performance. Chromium uses only one rect for partial invalidation.
• When window is resized, WR needs to render whole window.

Comment 2

[Sotaro Ikeda [:sotaro]]

:kamidphish, do you have an interest to this bug? It is also necessary for get performance with DirectComposition.

Comment 3

[Sotaro Ikeda [:sotaro]]

Attachment: temporal patch - Use IDXGISwapChain1::Present1 and WR dirty rects

The patch uses WR dirty rects(Bug 1581650 ) for present swap chain. But it does not expect to reduce GPU task with the current patch.

Comment 4

[Sotaro Ikeda [:sotaro]]

The following has an information about partial present.
https://docs.microsoft.com/en-us/windows/win32/direct3ddxgi/dxgi-1-2-presentation-improvements

Comment 5

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch - Use IDXGISwapChain1::Present1 and WR dirty rects

With the patch, black rectangle appeared during rendering update. Something seemed not work as expected.

Comment 6

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch - Use IDXGISwapChain1::Present1 and WR dirty rects

Comment 7

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch

Comment 8

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch

Comment 9

[Sotaro Ikeda [:sotaro]]

Attachment 9105152 [details] [diff] has2 problems.
-[1] CopySubresourceRegion() seemed not to work as expected

• But error did not reported with pref gfx.direct3d11.enable-debug-layer : true

-[2] Black square appeared when content showed a scrollable web page and mouse was around browser UI and content.

• The black square did not appear during showing about:support or about:config.

Comment 10

[Sotaro Ikeda [:sotaro]]

Attachment: temporal patch - Enable WR partial present and use one buffer swap chain

The patch enables partial present in WR. But it does full Present() at RenderCompositorANGLE::EndFrame(). It should work, since RenderCompositorANGLE allocates one buffer SwapChain with DXGI_SWAP_EFFECT_SEQUENTIAL. The SwapChain works as Bitblt model swap chain.
https://docs.microsoft.com/en-us/windows/win32/direct3ddxgi/dxgi-1-2-presentation-improvements#bitblt-model-swap-chain-with-dirty-rectangles

But with the patch, [2] in Comment 9 still happened. With it, WR's partial rendering might have a problem. The problem could happen more frequently with the followings.

• "Bookmarks Toolbar" is enabled and has links.
• Browser shows a web site.
• Move mouse around the links and the web site.

Comment 11

[Sotaro Ikeda [:sotaro]]

:gw, can you comment to Comment 10? Do you have an idea about how the black square could happen?

Comment 12

[Markus Stange [:mstange]]

I don't quite understand what you're doing here. Why is the CopySubresourceRegion call needed? According to the partial present documentation, Present1 should be doing all the necessary copying automatically. You just need to make sure to give it the right dirty rects and draw to those same dirty rects. (Maybe WebRender is drawing to the wrong dirty rects? Maybe the coordinates are upside down?)

Comment 13

[Sotaro Ikeda [:sotaro]]

You just need to make sure to give it the right dirty rects and draw to those same dirty rects.

Yea. You are right. I am going to check the coordinates. CopySubresourceRegion is necessary when WR uses partial present, but RenderCompositorANGLE does not do partial present.

Comment 14

[Markus Stange [:mstange]]

(In reply to Sotaro Ikeda [:sotaro] from comment #13)

CopySubresourceRegion is necessary when WR uses partial present,

Why?

Comment 15

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch

Removed CopySubresourceRegion usage.

Comment 16

[Sotaro Ikeda [:sotaro]]

(In reply to Markus Stange [:mstange] from comment #14)

(In reply to Sotaro Ikeda [:sotaro] from comment #13)

CopySubresourceRegion is necessary when WR uses partial present,

Why?

For example, in CompositorD3D11, partial present is disabled for NVIDIA GPU, but compositor renders only part of back buffer.

Comment 17

[Glenn Watson [:gw]]

(In reply to Sotaro Ikeda [:sotaro] from comment #16)

(In reply to Markus Stange [:mstange] from comment #14)

(In reply to Sotaro Ikeda [:sotaro] from comment #13)

CopySubresourceRegion is necessary when WR uses partial present,

Why?

For example, in CompositorD3D11, partial present is disabled for NVIDIA GPU, but compositor renders only part of back buffer.

In the case that partial present is disabled, could you initialize WR with max_partial_presents = 0, and then behavior is the same? Or is it likely faster to emulate partial present with the CopySubResourceRegion ?

Comment 18

[Sotaro Ikeda [:sotaro]]

(In reply to Glenn Watson [:gw] from comment #17)

In the case that partial present is disabled, could you initialize WR with max_partial_presents = 0, and then behavior is the same? Or is it likely faster to emulate partial present with the CopySubResourceRegion ?

It seems likely faster than simple full present.

Comment 19

[Sotaro Ikeda [:sotaro]]

I might find the cause of rendering problem. There might be 2 problems.

• [1] composite_simple() counts dirty_rect_count. But there is a case that dirty_rect_count == 0. In this case, WR does not do rendering. But RenderCompositorANGLE::EndFrame() flip swap chain, since no dirty rect is also used for meaning full rendering.
  • https://searchfox.org/mozilla-central/source/gfx/wr/webrender/src/renderer.rs#4010
• [2] PartialPresentMode::Multi seemed not work as expected. When PartialPresentMode::Multi is used, draw_tile_list() checks only dirty_rect that is set in a Tile. But there are overlapping Tiles. They also need to be re-rendered.
  • https://searchfox.org/mozilla-central/source/gfx/wr/webrender/src/renderer.rs#3926

Comment 20

[Sotaro Ikeda [:sotaro]]

:gw, can you comment to Comment 19 ?

Comment 21

[Markus Stange [:mstange]]

(In reply to Sotaro Ikeda [:sotaro] from comment #19)

I might find the cause of rendering problem. There might be 2 problems.

Nice work!

Comment 22

[Glenn Watson [:gw]]

Thanks for investigating, I'll take a look at this today!

Comment 23

[Sotaro Ikeda [:sotaro]]

Thanks!

Comment 24

[Glenn Watson [:gw]]

Attachment: 0001-WIP.patch

There is a patch up to fix the partial present semantics in WR (https://bugzilla.mozilla.org/show_bug.cgi?id=1593154).

This WIP patch builds on top of this - it skips Present1() call completely if WR returns no dirty rects at all.

From what I can see, this fixes the issues with partial present. I didn't notice any glitches during browsing. I also saw from logging that we are often presenting very small rects, or none at all (completely skipping Present1). It also fixes the error code returns we were getting from Present1() when WR was supplying an empty dirty rect.

Comment 25

[Sotaro Ikeda [:sotaro]]

Attachment: Bug 1575159 - Implement partial invalidation on Windows

Comment 26

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/1e151275792d
Implement partial invalidation on Windows r=gw

Comment 27

[Daniel Varga [:dvarga]]

https://hg.mozilla.org/mozilla-central/rev/1e151275792d