With this CL, SimulcastEncoderAdapter no longer configures its encoder
as multi-layered if we only have a single active layer. Instead we
create a single single-layered encoder for that one and only active
layer. When using VP8 SW encoder this means that LibvpxVp8Encoder is
configured to only prepare a single video frame which avoids the cost of
scaling down to layers that we do not send. (A multi-layered
LibvpxVp8Encoder is required to scale even layers we don't encode.)
When profiling this CL I found very small but measurable gains for
representative downscale factors of 20.1 ms of 60 s profile. This is
just 0.0335% CPU so it's not much, but skipping a downscale might be
worth a lot more if we have to map/unmap buffers or do GPU round-trips
in the future (which I have not measured).
When downscaling to factors 4 and 2 due to libyuv having a
"fast-path" for these (i.e. no adaptation active), zero difference was
found for NV12. For I420 there was small regression of 16.1 ms
(0.026% CPU) for this one edge-case. It's possible to work around this,
but considering the tiny changes we're talking about, I really don't
think it's worth the additional complexity. I'll file a bug on libyuv
about scaling factors 2+2 vs 4 and leave it at that.
Bug: webrtc:12603
Change-Id: Id462140c6a829cf6b460baae868e94243f477db3
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/219683
Commit-Queue: Henrik Boström <hbos@webrtc.org>
Reviewed-by: Ilya Nikolaevskiy <ilnik@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#34092}
70cd086644