commit 4be5927dc788b814dff2a6ca46175254dce59232
author henrika <henrika@webrtc.org> Thu Nov 02 09:16:00 2023
committer WebRTC LUCI CQ <webrtc-scoped@luci-project-accounts.iam.gserviceaccount.com> Fri Nov 03 18:05:17 2023
tree ce6fd61e7a453e567050e2a4eeb906deca78221b
parent 2931ddd2e991d31958a1f59066ecb8c5923af892

Reduces rate at which TryGetNextFrame returns NULL for WGC

This CL is a follow-up of work done in
https://webrtc-review.googlesource.com/c/src/+/323882 where the goal
was to reduce the amount of FrameDropped error logs in
WebRTC.DesktopCapture.Win.WgcCaptureSessionGetFrameResult.

The previous work avoids FrameDropped logs for a minimized window
being captured with WGC but we still se a large amount of these error
(or rather warning) logs. See [1] which comes from Canary.

This CL does two different things to improve the situation:

1) It adds kFramePoolEmpty to the existing
GetFrameResult::kFrameDropped enum to point out that the warning
comes from the frame pool not being able to return a valid new frame.
It also makes it more clear that it does not cause an outer/final
error as WgcCapturerResult::kFrameDropped. We still keep the inner
GetFrameResult::kFrameDropped but it is only produced when the frame
pool returns NULL and our external queue is empty. Hence, a real
frame-drop error. Note that, it is still easy to provoke
kFramePoolEmpty simply by asking for a high resolution at a high rate.
The example in [2] comes from a 4K screen @30fps. Hence, we have not
fixed anything yet.

2) It also increases the size of the internal frame pool from 1 to 2.
This does lead to an almost zero rate of kFramePoolEmpt
warnings at the expense of a slightly reduced max capture rate. BUT,
with 1 as size, we can "see" a higher max capture rate but it is not
a true rate since it comes with a high rate of kFramePoolEmpty
errors. Hence, we "emulate" a high capture rate by simply feeding
copies of the last frame that we had stored in the external queue.
Using 2 leads to a more "true" rate of what we actually can capture
in terms of *new* frames and also a substantially lower rate of
kFramePoolEmpty.
In addition, with 1 as size, if we ask at a too high rate and provide
a copy of the last frame, our CPU adaptation will not reduce its rate
since we think that things are OK when it is actually not.

Also, the samples in [3] and [4] both use 2 as numberOfBuffers
as well.

Let me also mention that with this small change, I a have not been
able to provoke any kFramePoolEmpty error messages.

Finally, geDisplayMedia can be called called with constraints where
min and max framerate is defined. The mechanism which maintains the
min rate is implemented via the RequestRefreshFrame API and it can
be sent to the source (DesktopCaptureDevice) back to back with a
previous timer interrupt for a capture request. Without this change,
these RRFs were also a source of a large amount of
kFramePoolEmpty error logs. With 2 buffers instead; this is no
longer the case.

[1] https://screenshot.googleplex.com/7sfv6HdGXLwyxdj
[2] https://paste.googleplex.com/4795680001359872
[3] https://github.com/robmikh/Win32CaptureSample/blob/master/Win32CaptureSample/SimpleCapture.cpp
[4] https://learn.microsoft.com/en-us/windows/uwp/audio-video-camera/screen-capture#add-the-screen-capture-capability

Bug: chromium:1314868
Change-Id: I73b823b31a993fd2cd6e007b212826dfe1a80012
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/325521
Commit-Queue: Alexander Cooper <alcooper@chromium.org>
Reviewed-by: Alexander Cooper <alcooper@chromium.org>
Cr-Commit-Position: refs/heads/main@{#41079}