search

mozilla / pdf.js

PDF Reader in JavaScript
https://mozilla.github.io/pdf.js/
Apache License 2.0
48.16k stars 9.95k forks source link

Add an option to enable/disable hardware acceleration (bug 1902012) #18238

Closed calixteman closed 3 months ago

calixteman commented 3 months ago

/botio test

moz-tools-bot commented 3 months ago

From: Bot.io (Linux m4)

Received

Command cmd_test from @calixteman received. Current queue size: 0

Live output at: http://54.241.84.105:8877/9f3ddb12d4070bc/output.txt

moz-tools-bot commented 3 months ago

From: Bot.io (Windows)

Received

Command cmd_test from @calixteman received. Current queue size: 0

Live output at: http://54.193.163.58:8877/caa82771b7db49d/output.txt

moz-tools-bot commented 3 months ago

From: Bot.io (Linux m4)

Failed

Full output at http://54.241.84.105:8877/9f3ddb12d4070bc/output.txt

Total script time: 28.87 mins

  different ref/snapshot: 16
  different first/second rendering: 2

Image differences available at: http://54.241.84.105:8877/9f3ddb12d4070bc/reftest-analyzer.html#web=eq.log

moz-tools-bot commented 3 months ago

From: Bot.io (Windows)

Failed

Full output at http://54.193.163.58:8877/caa82771b7db49d/output.txt

Total script time: 41.61 mins

  different ref/snapshot: 5

Image differences available at: http://54.193.163.58:8877/caa82771b7db49d/reftest-analyzer.html#web=eq.log

calixteman commented 3 months ago

/botio test

moz-tools-bot commented 3 months ago

From: Bot.io (Linux m4)

Received

Command cmd_test from @calixteman received. Current queue size: 0

Live output at: http://54.241.84.105:8877/95003cd5491e9e1/output.txt

moz-tools-bot commented 3 months ago

From: Bot.io (Windows)

Received

Command cmd_test from @calixteman received. Current queue size: 0

Live output at: http://54.193.163.58:8877/c03d7a10b723af1/output.txt

moz-tools-bot commented 3 months ago

From: Bot.io (Linux m4)

Failed

Full output at http://54.241.84.105:8877/95003cd5491e9e1/output.txt

Total script time: 29.04 mins

  different ref/snapshot: 16
  different first/second rendering: 2

Image differences available at: http://54.241.84.105:8877/95003cd5491e9e1/reftest-analyzer.html#web=eq.log

moz-tools-bot commented 3 months ago

From: Bot.io (Windows)

Failed

Full output at http://54.193.163.58:8877/c03d7a10b723af1/output.txt

Total script time: 44.19 mins

  different ref/snapshot: 5

Image differences available at: http://54.193.163.58:8877/c03d7a10b723af1/reftest-analyzer.html#web=eq.log

calixteman commented 3 months ago

/botio test

moz-tools-bot commented 3 months ago

From: Bot.io (Linux m4)

Received

Command cmd_test from @calixteman received. Current queue size: 0

Live output at: http://54.241.84.105:8877/26467fb15f49b35/output.txt

moz-tools-bot commented 3 months ago

From: Bot.io (Windows)

Received

Command cmd_test from @calixteman received. Current queue size: 0

Live output at: http://54.193.163.58:8877/e164b4315fc63ec/output.txt

moz-tools-bot commented 3 months ago

From: Bot.io (Linux m4)

Failed

Full output at http://54.241.84.105:8877/26467fb15f49b35/output.txt

Total script time: 28.62 mins

  different ref/snapshot: 16
  different first/second rendering: 2

Image differences available at: http://54.241.84.105:8877/26467fb15f49b35/reftest-analyzer.html#web=eq.log

moz-tools-bot commented 3 months ago

From: Bot.io (Windows)

Failed

Full output at http://54.193.163.58:8877/e164b4315fc63ec/output.txt

Total script time: 43.71 mins

  different ref/snapshot: 8

Image differences available at: http://54.193.163.58:8877/e164b4315fc63ec/reftest-analyzer.html#web=eq.log

qwer1304 commented 3 months ago

Why isn't the option exposed in Chrome's options? ("title": "Enable HW acceleration" is missing below line 48 in extensions/chromium/preferences_schema.json)

Rob--W commented 3 months ago

While disabling HWA unconditionally will certainly fix some issues associated with https://bugzilla.mozilla.org/show_bug.cgi?id=1902012, I wonder whether doing this unconditionally by default is the correct approach to take.

Do we have any data that shows that turning off HWA unconditionally is going to have a larger positive impact than the performance loss on systems where HWA outperforms software rendering?

Browsers usually have hardware blocklists to selectively disable HWA on systems that are known to have issues with HWA. Additionally, users can override the browser defaults, chrome://flags/#disable-accelerated-2d-canvas in Chrome, or gfx.canvas.accelerated=false in about:config in Firefox.

This patch does not only add an option to disable HWA, it already disables HWA in situations where the PDF.js code is expected to read back data, by specifying willReadFrequently:true unconditionally.

calixteman commented 3 months ago

@Rob--W, I wasn't so happy to disable hwa for the exact reasons you gave (and tbh I was pretty reluctant). We've now some performance tests in m-c and I hope we'll have some numbers soon. An other thing we could do is to create a Nimbus experiment to try measure if there are significant difference when rendering the first page for the different cohorts.

Rob--W commented 3 months ago

If there are doubts about disabling unconditionally, could we change the default to true (so that the use of HWA is controlled by the browser/user)? The mere existence of the setting now enables others to opt out if they want to deploy PDF.js in an environment where HWA is an issue.

PDF.js + HWA has been out in the wild for many years at this point, whereas HWA off is a new change here. I think that it would be lower risk to default to true (since it is the same as before), and then experiment with turning it off in cohorts. If that shows a noticeable performance difference, then the default can change.

marco-c commented 3 months ago

IIRC canvas acceleration has only been available recently in Firefox, so effectively it's only a short time HWA has been really enabled in Firefox. I think we can quickly estimate the effects by running the perf tests and comparing, this will give us a good initial idea.

Rob--W commented 3 months ago

It's indeed relatively new, but not trivial either (1.5 years).

HWA canvas has been enabled on Linux+macOS since Firefox 110 (https://bugzilla.mozilla.org/show_bug.cgi?id=1806058), on Android since Firefox 113 (https://bugzilla.mozilla.org/show_bug.cgi?id=1825182). Interestingly, canvas HWA seems not enabled on Windows: https://searchfox.org/mozilla-central/rev/94f839e924ba6c69f3e0d062d4c6cc4fee7cad5b/modules/libpref/init/StaticPrefList.yaml#5706-5715

In Chrome, HWA canvas has been around for 12 years, judging by the announcement at https://developer.chrome.com/blog/taking-advantage-of-gpu-acceleration-in-the-2d-canvas/

If there is a desire to turn off HWA by default, I'd like it to be enabled by default in the Chrome extension (I can submit a patch), unless there is data that shows that HWA off-by-default offers a better experience.

I think we can quickly estimate the effects by running the perf tests and comparing, this will give us a good initial idea.

Is this only in Firefox, or does it also cover other browser engines such as Chromium?

qwer1304 commented 3 months ago

At least I can tell you that I've been having issues when HWA was turned OFF in pdf.js running in Chrome as listed in issue 18242. That's why I asked about making it a user controlled option in Chrome. After enabling it, the weird issues have gone away (for the meantime). It's worth mentioning that it seems that the pdf.js option is NOT completely equivalent to the Chrome's flag option (chrome://flags/#disable-accelerated-2d-canvas), since I've been having issues even with the HWA disabled in Chrome whilst it was disabled in pdf.js. Only after enabling HWA in pdf.js did the issues gone away.