VP9 Encoding

Converting videos to VP9 is a fantastic solution for those that want to compress videos to save bandwidth/storage without losing perceptual quality, and be able to play those videos back in the browser (e.g. a video site like youtube). In learning to use VP9, I have had a lot of poor results along the way. With this codec, it really pays to look into the parameters and specify what you want for your use case.

Related Posts

• Comparing VP9 AND x265 HEVC

Conversion Script

Below is the BASH script I use to encode my videos to VP9 webm after having installed ffmpeg through a PPA. I find that this script is a good balance between encoding speed and quality and I explain the parameters below it. It is a 2-pass encode targeting an overall bitrate.

#!/bin/bash
EXPECTED_NUM_ARGS=3;

if [ "$#" -ne $EXPECTED_NUM_ARGS ]; then
    echo "Expecting 3 arguments: [input filename] [output filename] [bitrate in K]"
    exit 1
fi

INPUT_FILE=$1
OUTPUT_FILE=$2
BITRATE=$3
BITRATE="`echo $BITRATE`K"

EXPECTED_NUM_ARGS=3;

if [ "$#" -ne $EXPECTED_NUM_ARGS ]; then
    echo "Illegal number of aguments"
fi

INPUT_FILE=$1
OUTPUT_FILE=$2
BITRATE=$3
BITRATE="`echo $BITRATE`K"
NUM_CORES=$(cat /proc/cpuinfo | grep processor | wc -l)

ffmpeg \
  -i $INPUT_FILE \
  -c:v libvpx-vp9 \
  -pass 1 \
  -b:v $BITRATE \
  -g 150 \
  -threads $NUM_CORES \
  -speed 4 \
  -tile-columns 4 \
  -frame-parallel 0 \
  -an -f webm /dev/null \
  && \
    ffmpeg \
    -i $INPUT_FILE \
    -c:v libvpx-vp9 \
    -pass 2 \
    -b:v $BITRATE \
    -g 150 \
    -threads $NUM_CORES \
    -speed 1 \
    -tile-columns 4 \
    -frame-parallel 0 \
    -auto-alt-ref 1 \
    -lag-in-frames 25 \
    -auto-alt-ref 1 \
    -c:a libopus \
    -b:a 128k \
    -f webm \
    $OUTPUT_FILE

• -g 150 (or kf_max_dist)- There will be a maximum of 150 frames between keyframes. If you don't set this, it will be an absurdly high number 99999 and you will probably notice that when you seek around in a video will be slow.
• -threads 4 - Multi-threaded encoding may be used if -threads > 1 and -tile-columns > 0.
  • It used to be the case that a 1920 pixel wide video could only use 4 tile-columns and thus 4 threads, but I believe now you can use any number of threads you like.
• -tile-columns 4 - using tile-columns slightly reduces quality. However it improves encoding and decoding speed by allowing multiple threads to be used. Most of my content has a 1920 pixel width so the maximum I can have is 4 with 4 threads.
• -frame-parallel - Looks like this was once required for multithreading but no more so best to turn it off.
• -cpu-used or -speed - sets how efficient the compression will be, trading faster encoding for a larger file size.
  • Generally good to use high speed for the first pass, and slow speed for the second pass.
  • For legacy reasons, -speed is an alias for -cpu-used.
• Setting auto-alt-ref and lag-in-frames >= 12 will turn on VP9's alt-ref frames, a VP9 feature that enhances quality. More info here.
• -row-mt 1 enables row-based multi-threading which maximizes CPU usage. Enabling row-mt is only faster when the CPU has more threads than the number of encoded tiles.

References

• FFMPEG Wiki: Encode VP9
• Webm Wiki - FFmpeg‎ > ‎VP9 Encoding Guide
• Github - Kagami/webm.py - Notes on encoding settings
• Webm - VP8 Encode Parameter Guide
• Stack Overflow - VP9 encoding limited to 4 threads?
• Phoronix - VP9 Encoder Gets Better Multi-Threading Performance
• FFmpeg - AV1 Video Encoding Guide