Tweaking OSR performance in Kivy example

[cztomczak]

windowless_frame_rate

In CEF 51 there is a new option "windowless_frame_rate" in BrowserSettings.

---

Chromium flags

There are a bunch of flags in Chromium that can be used to tweak OSR, I see people posting them from time to time on the CEF Forum and in the CEF Issue Tracker.

You can pass flags to CEF Python programmatically, see api/CommandLineSwitches.md.

Try for example combinations of these flags:

--disable-gpu --disable-gpu-compositing
--enable-begin-frame-scheduling
--disable-d3d11

For best performance pass --disable-gpu --disable-gpu-compositing --enable-begin-frame-scheduling - info by Marshall from the CEF Forum. These three flags are added by default in upstream cefclient OSR mode. Note that this will disable WebGL.

--disable-gpu --disable-gpu-compositing

When the disable-gpu and disable-gpu-compositing command-line flags are specified the CefRenderHandler::OnPaint method will be called directly from the compositor instead of requiring an additional copy for each frame.

Use software rendering and compositing (disable GPU) for increased FPS and decreased CPU usage. This will also disable WebGL so remove these switches if you need that capability.See https://bitbucket.org/chromiumembedded/cef/issues/1257 for details.

These flags are passed by default when calling cefclient with OSR enabled. If you pass --enable-gpu to cefclient then these flags will not be added.

--enable-begin-frame-scheduling

Synchronize the frame rate between all processes. This results in decreased CPU usage by avoiding the generation of extra frames that would otherwise be discarded. The frame rate can be set at browser creation time via CefBrowserSettings.windowless_frame_rate or changed dynamically using CefBrowserHost::SetWindowlessFrameRate. In cefclient it can be set via the command-line using --off-screen-frame-rate=XX. See https://bitbucket.org/chromiumembedded/cef/issues/1368 for details.

--disable-d3d11

Windows-only to disable use of Direct3D 11.

---

cefclient

You can compare performance with the cefclient sample application from upstream. The cefclient executable is distributed along with CEF Python binaries. To run it in OSR mode type:

./cefclient --off-screen-rendering-enabled

By default frame rate is 30, to change it:

./cefclient --off-screen-rendering-enabled --off-screen-frame-rate=60

Additional flags that can be passed to cefclient:

--off-screen-frame-rate
--enable-gpu
--transparent-painting-enabled
--multi-threaded-message-loop
--external-message-pump

When --enable-gpu is not passed, then cefclient adds these flags automatically by default: --disable-gpu --disable-gpu-compositing

Up-to-date OSR performance related flags can be found in upstream cefclient > ClientAppBrowser::OnBeforeCommandLineProcessing.

---

Message loop work

There is this line in kivy_.py to run message loop work:

#start idle
Clock.schedule_interval(self._cef_mes, 0)

Is this the best optimized way? The comment states "idle", which makes me wonder. Here is the reference in Kivy docs: https://kivy.org/docs/api-kivy.clock.html

Clock.schedule_once(my_callback, 0) # call after the next frame
Clock.schedule_once(my_callback, -1) # call before the next frame

How can we call message loop work in Kivy using constant intervals like 10ms?

---

PDF rendering in OSR mode

Not really performance related, but will mention it here for now. If the PDF extension is enabled then cc Surfaces must be disabled for PDFs to render correctly. For details see CEF Issue 1689:
https://bitbucket.org/chromiumembedded/cef/issues/1689

So:

1. Either disable PDF extension using these flags: --disable-extensions --disable-pdf-extension
2. Or if PDF extension is enabled then in OSR mode you need to pass the --disable-surfaces flag

Is --disable-surfaces flag deprecated? Can't find it neither using cs.chromium.org nor peter.sh flags

[jegger]

This is really tricky to debug because I can't find a way to measure the "real" frame-reate. Kivy's frame-rate is always around 70-80 (python kivy_.py -m monitor). I found https://frames-per-second.appspot.com which seems helpful. You get the framerate by clicking the green/yellow button in the right top corner (Browser Paint). - But is this really the fps you see?

Without any new command line switches / settings:
I get 30fps browser-paint maximum. The ball (60fps) is (very) laggy.

With only windowless_frame_rate set:
I get 60fps browser-paint. But the browser-paint fps drops down to 55-60 from time to time. The ball (60fps) seems laggier than the ball in kivy+old cefpython (which also has 60fps browser-paint).

With only --off-screen-frame-rate=60:
I get 30fps browser-paint. Does this really work? ( I see getting it passed to zygote, so I guess it is not a typo of mine)

With windowless_frame_rate set and --disable-gpu-compositing:
I get 60fps browser-paint. And no more drops below 60.
(disable-gpu and enable-begin-frame-scheduling does not seem to take any major effects)
I can't say that it is now the same "speed and feel" as with old cefpython+kivy. I will try to make a slow-motion capture of the screen to find that out.

btw --disable-gpu-vsync:
Lets the browser-paint fps go up to 80-90. But I don't know if this is any good?

[cztomczak]

Improved performance in the Kivy example in commit 638507a:

• Use the same Chromium flags as in upstream cefclient
• Call message loop work in faster 6ms intervals by using a small trick. By default when scheduling after the next frame intervals are 13ms.

[jegger]

I guess this is now pretty good.

But in general, could it be a vsync issue? If we have "multiple" loops could that be an issue?
I never looked into that (not even in kivy itself).
Or does CEF render at 60hz anyways? And MessageLoopWork() just "fetches" the newest texture.

[cztomczak]

I can only suggest asking these questions on the CEF Forum.

Recently an external message pump functionality was added to CEF that can improve performance. Created Issue #246 for that.

[allestuetsmerweh]

Note: To measure the real browser FPS, one could check how often the framebuffer from CEF changes.
This could be done here:
https://github.com/allestuetsmerweh/garden.cefpython/blob/master/cefbrowser/cefbrowser.py#L882

[allestuetsmerweh]

We now have an FPS meter included in garden.cefpython (kivy), when run with --debug:
allestuetsmerweh/garden.cefpython@f953ea6

[cztomczak]

There is Issue #2046 ("Improve GPU readback performance for OSR") in upstream CEF along with PR #96 submitted, which might increase OSR performance significantly.

[cztomczak]

Added FPS meter in commit 2516613 (pass "fps" arg).

I was testing Application.external_message_pump in latest v58 and some things got better, but there are still issues. Loading time of web pages is now fixed and behaves the same as in default loop. See these last two comments for details: #246 (comment)