1352163 - Use double buffering for D3D11 window backbuffers instead of blitting

Status: RESOLVED FIXED

(Core :: Graphics: Layers, enhancement)

Description

[Bas Schouten (:bas.schouten)]

Right now we blit on each present. On large changes this consumes quite a lot of bandwidth on higher DPI screens. We should use a buffer flip model and save that bandwidth.

Comment 1

[Bas Schouten (:bas.schouten)]

Attachment: Bug 1352163: Use buffer flipping for presenting to the browser window.

Review commit: https://reviewboard.mozilla.org/r/125162/diff/#index_header
See other reviews: https://reviewboard.mozilla.org/r/125162/

Comment 2

[Matt Woodrow (:mattwoodrow)]

Comment on attachment 8853075 [details]
Bug 1352163: Use buffer flipping for presenting to the browser window.

https://reviewboard.mozilla.org/r/125162/#review127760

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/1765f0af5161
Use buffer flipping for presenting to the browser window. r=mattwoodrow

Comment 4

[Carsten Book [:Tomcat]]

https://hg.mozilla.org/mozilla-central/rev/1765f0af5161

Comment 5

[Joel Maher ( :jmaher ) (UTC -8)]

this has a perf win on windows 7 where we measure fileio for startup:
== Change summary for alert #5819 (as of April 03 2017 17:23 UTC) ==

Improvements:

 23%  tp5n nonmain_startup_fileio windows7-32 opt e10s     54181.12 -> 41804.08

For up to date results, see: https://treeherder.mozilla.org/perf.html#/alerts?id=5819

Comment 6

[[SV Manager] Florin Mezei, QA (:FlorinMezei)]

Bas, what are the testing needs for this change? Any sort of manual/automated testing needed for this?

Same question related to bug 1352442.

Comment 7

[u279076]

(In reply to Florin Mezei, QA (:FlorinMezei) from comment #6)
> Bas, what are the testing needs for this change? Any sort of manual/automated testing needed for this?

Hi Florin,

I won't remove the need-info for Bas since he probably has some more info but I wanted to let you know that I'm planning a Telemetry Experiment to A/B test this on Nightly. If I'm understanding this correctly, on certain systems (Win10 w/HiDPI screens) we'd expect this to improve D3D performance without regressing stability. My experiment will measure this to see what evidence we can gather on that assumption.

In terms of manual testing the only thing I would suggest is standing up some systems which meet the hardware/platform requirements and then test GPU accelerated web content with both different display configurations (High DPI, Low DPI, multiple screens with different DPIs, hotswapping screens, etc).

Bas can likely answer the question on what automation exists to help vet this further.