Noticeable speed regression since v1.6.1

[Kagami Hiiragi]

Hi.

While encoding videos with different versions of libvpx, I noticed that
newer versions are slower, especially since v1.6.1. Here is simple
script to test:

#!/bin/bash
set -eu
export TIMEFORMAT="%P%% cpu %R total"
VS=( v1.5.0 v1.6.0 v1.6.1 master )
for V in ${VS[@]}; do
echo "# $V"
git checkout -q $V
git clean -q -xdf
./configure --disable-unit-tests --disable-docs &>/dev/null
make -j8 &>/dev/null
time ./vpxenc -q --codec=vp9 --passes=1 --good --cpu-used=0 \
--end-usage=q --cq-level=30 -o /dev/null ~/blue_sky_360p_60f.y4m
done

Result:

$ ../test-libvpx.sh
# v1.5.0
125.27% cpu 23.702 total
# v1.6.0
124.87% cpu 26.408 total
# v1.6.1
125.35% cpu 30.609 total
# master
125.49% cpu 30.627 total

This is only single video, but I can reproduce it on anything. As you
can see, regression is quite significant, about 1.3x slowdown comparing
v1.5.0 and master. (My CPU is i7 3820 and I'm on 64bit Linux if that
matters.)

So my question: is that expected behavior and why does it happen? Is it
because later versions give you better quality, justifying slowdown? Or
some other reason?

Thanks.

[Szymon Pałka]