Improve CPU utilization when encoding VP8 with a single temporal layer
Prior to this CL, when software VP8 encoding was done with one temporal
layer, instead of only predicting from the latest frame, the code
allowed the encoder to reference the latest key frame as well.
This improves quality for the few frames immediately after
the key frame, but is not useful for later frames, which diverge
significantly from the key frame. However, the cost of producing
the prediction from more than one reference is incurred by all frames.
My measurements of the effect of this show an improvement
in CPU utilization of 5%-13% when this is not done.
foreman_352x288, 30fps, target bitrate 500kps
Pre-change:
send_enc_speed_fps: avg(566.187, 570.012, 575.665) = 570.621
send_avg_qp: 45.36
send_avg_psnr: 37.13
Post-change:
send_enc_speed_fps: avg(633.188, 604.694, 623.232) = 620.371
send_avg_qp: 45.88
send_avg_psnr: 37.0749
Improvement in send_enc_speed_fps: 8.71%
foreman_480x272, 30fps, target bitrate 500kps
Pre-change:
send_enc_speed_fps: avg(481.244, 486.971, 487.322) = 485.179
send_avg_qp: 48.9
send_avg_psnr: 37.6217
Post-change:
send_enc_speed_fps: avg(521.651, 499.416, 511.551) = 510.872
send_avg_qp: 48.88
send_avg_psnr: 37.6094
Improvement in send_enc_speed_fps: 5.29%
news_352x288, 30fps, target bitrate 500kps
Pre-change:
send_enc_speed_fps: avg(699.407, 697.837, 699.49) = 698.9113333
send_avg_qp: 24.15
send_avg_psnr: 40.9551
Post-change:
send_enc_speed_fps: avg(758.526, 768.104, 757.232) = 761.2873333
send_avg_qp: 23.9833
send_avg_psnr: 40.9697
Improvement in send_enc_speed_fps: 8.92%
Bridge_180x320_15 (video of brandtr@ from Google), 15fps, target bitrate 500kps
Pre-change:
send_enc_speed_fps: avg(454.757, 450.399, 446.812) = 450.656
send_avg_qp: 17.6771
send_avg_psnr: 39.9267
Post-change:
send_enc_speed_fps: avg(500.014, 513.316, 513.613) = 508.981
send_avg_qp: 17.6837
send_avg_psnr: 39.9137
Improvement in send_enc_speed_fps: 12.94%
Bug: webrtc:10281
Change-Id: If02736e1535c5f46689fd42b657e35a1e1f64d6d
Reviewed-on: https://webrtc-review.googlesource.com/c/120904
Reviewed-by: Sergey Silkin <ssilkin@webrtc.org>
Reviewed-by: Stefan Holmer <stefan@webrtc.org>
Commit-Queue: Elad Alon <eladalon@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#26511}diff --git a/modules/video_coding/codecs/vp8/default_temporal_layers.cc b/modules/video_coding/codecs/vp8/default_temporal_layers.cc
index ee06bfa..674e7d9 100644
--- a/modules/video_coding/codecs/vp8/default_temporal_layers.cc
+++ b/modules/video_coding/codecs/vp8/default_temporal_layers.cc
@@ -134,8 +134,8 @@
// keyframe.
switch (num_layers) {
case 1:
- // All frames reference all buffers and the 'last' buffer is updated.
- return {Vp8FrameConfig(kReferenceAndUpdate, kReference, kReference)};
+ // Always reference and update the same buffer.
+ return {Vp8FrameConfig(kReferenceAndUpdate, kNone, kNone)};
case 2:
// All layers can reference but not update the 'alt' buffer, this means
// that the 'alt' buffer reference is effectively the last keyframe.
