tree b36e1bb2ab7ddc6b3f32b980a83d73703567a40d
parent 2c893e61ff47f09b763be48d6a78a20e653eabaa
author Hui Su <huisu@google.com> 1533861670 -0700
committer Hui Su <huisu@google.com> 1534206075 +0000

Add an ML based partition search breakout feature

Use a neural net model to predict if the encoder should breakout from
partition search after trying PARTITION_NONE. This patch only
implements models for small resolutions(less than 720P).

This feature is enabled for speed 0 and above. Coding loss:
speed 0:
lowres 0.088%   midres 0.073%   objective-1-fast(360P) -0.07%
speed 1:
lowres 0.051%   midres 0.080%   objective-1-fast(360P) 0.01%

Tested encoding speed over 20 lowres sequences, speed gains:
Speed 0:
             QP=35        QP=45       QP=55
average      11.39%      14.65%       20.56%
maximum      29.95%      53.47%       45.93%

Speed 1:
             QP=35        QP=45       QP=55
average       7.52%       7.38%       9.86%
maximum      18.20%      13.44%      25.85%

STATS_CHANGED

Change-Id: Iafff5da49db6ee7d304f69b0cd278160a78cb01f