Intel GPUs have dedicated hardware that enables fast, energy-efficient encoding and decoding of video compliant with the industry standards such as H.264, HEVC and AV1. Drivers and software enable application developers to utilize this hardware either directly from their own applications or from within popular frameworks such as FFmpeg and GStreamer.
This repository curates a coherent set of software, drivers and utilities to assist users who are getting started with transcoding video on Intel GPUs so that they can begin evaluating performance and incorporating Intel GPU transcode acceleration into their applications.
Included here are scripts, Docker configurations and documents that provide:
- Kernel mode drivers.
- The latest Intel GPUs are not fully supported by currently available Linux distributions so backported drivers can be installed by following the instructions below.
- User mode drivers and libraries:
- Transcoding software:
- FFmpeg with ffmpeg-qsv plugins (qsv means "QuickSync", a name for Intel's GPU video transcoding technology)
- Quality and performance measuring software:
- Set of quality and performance benchmarking scripts
- VMAF
- Intel GPU Tools
- Demonstrations:
- Samples which demonstrate operations typical for Content Delivery Network (CDN) applications such as Video On Demand (VOD) streaming under Nginx server.
- Setup guides and user documentation:
- GPU virtualization setup guide_
- Capabilities of ffmpeg-qsv plugins
- Benchmarking data for specific GPUs:
We recommend to attach GPU of Intel® Data Center GPU Flex Series to the system with the following characteristics:
| Recommendation | |
|---|---|
| CPU | 3rd Generation Intel® Xeon® Scalable Processors |
| Storage | At least of 20GB of free disk space available for docker |
| Operating System(s) | Ubuntu 20.04 or Ubuntu 22.04 with kernel mode support as described here |
GPU of Intel® Arc™ A-Series Graphics should be attached to the system with the following requirements:
| Recommendation | |
|---|---|
| CPU | One of the following CPUs:
|
| Storage | At least of 20GB of free disk space available for docker |
| Operating System(s) | Ubuntu 20.04 or Ubuntu 22.04 with kernel mode support as described here |
To run upstream GPU products, i.e. those GPUs support for which is available in Linux distributions, you need:
| Recommendation | |
|---|---|
| Operating System(s) | Linux distribution capable to support your GPU. We recommend Ubuntu 20.04 or later. |
| Storage | At least of 20GB of free disk space available for docker |
We provide dockerfiles to setup transcoding environment. These dockerfiles need to be built into docker containers. Dockerfiles can also be used as a reference instruction to install ingredients on bare metal.
If you want to run in virtual environment, first follow instruction given in GPU virtualization setup guide_ and then proceed with the setup described below.
Docker is required to build and to run media delivery containers. If you run Ubuntu 20.04 or later you can install it as follows:
sudo apt-get install docker.ioYou might need to further configure docker on your system:
Allow docker to run under your user account (remember to relogin for group modification to take effect):
sudo usermod -aG docker $(whoami) && exitConsider to register and login to Docker Hub. Docker Hub limits the number of docker image downloads ("pulls"). For anonymous users this limit is tied to IP address. For authenticated users it depends on subscription type. For details see https://docs.docker.com/docker-hub/download-rate-limit/. To authenticate running docker engine, execute:
# you will be prompted to enter username and password: docker loginIf you run behind a proxy, configure proxies for for docker daemon. Refer to https://docs.docker.com/config/daemon/systemd/. Below example assumes that you have
http_proxyenvironment variable set in advance:sudo mkdir -p /etc/systemd/system/docker.service.d echo "[Service]" | sudo tee /etc/systemd/system/docker.service.d/https-proxy.conf echo "Environment=\"HTTPS_PROXY=$http_proxy/\"" | \ sudo tee -a /etc/systemd/system/docker.service.d/https-proxy.conf sudo systemctl daemon-reload sudo systemctl restart docker sudo systemctl show --property=Environment --no-pager dockerMake sure that docker has at least 20GB of hard disk space to use. To check available space run (in the example below 39GB are available):
$ df -h $(docker info -f '{{ .DockerRootDir}}') Filesystem Size Used Avail Use% Mounted on /dev/sda1 74G 32G 39G 46% /If disk space is not enough (for example, default
/var/lib/dockeris mounted to a small size partition which might be a case for/var), consider reconfiguring docker storage location as follows:# Below assumes unaltered default docker installation when # /etc/docker/daemon.json does not exist echo "{\"data-root\": \"/mnt/newlocation\"}" | sudo tee /etc/docker/daemon.json sudo systemctl daemon-reload sudo systemctl restart docker
We provide few different setup configurations which differ by versions and origins of the included Intel media stack components. Some versions of media stack require special setup for the host.
| Dockerfile | Intel media stack origin | Supported Intel GPUs | Host setup instructions |
|---|---|---|---|
| docker/ubuntu20.04/selfbuild-prodkmd/Dockerfile | Self-built from open source | Alchemist, ATS-M | Intel GPU DKMS |
| docker/ubuntu20.04/selfbuild/Dockerfile | Self-built from open source | Gen8+ (legacy upstreamed platforms), such as SKL, KBL, CFL, TGL, DG1, etc. | Use any Linux distribution which supports required platform |
| docker/ubuntu20.04/native/Dockerfile | Ubuntu 20.04 | Gen8+, check Ubuntu 20.04 documentation | Use any Linux distribution which supports required platform |
To build any of the configurations, first clone Media Delivery repository:
git clone https://github.com/intel/media-delivery.git && cd media-deliveryTo build configuration which targets DG2/ATS-M stack self-built from open source projects, run:
docker build \
$(env | grep -E '(_proxy=|_PROXY)' | sed 's/^/--build-arg /') \
--file docker/ubuntu20.04/selfbuild-prodkmd/Dockerfile \
--tag intel-media-delivery .To build configuration which targets Gen8+ legacy upstreamed platforms via stack self-built from open source projects, run:
docker build \
$(env | grep -E '(_proxy=|_PROXY)' | sed 's/^/--build-arg /') \
--file docker/ubuntu20.04/selfbuild/Dockerfile \
--tag intel-media-delivery .Docker containers provide isolated environments with configured software. To access resources on a host system you need to add specific options when starting docker containers. Overall, software included into media-delivery constainers requires the following:
- To access desired GPU you need to map it to the container, see
--deviceoption below - To be able to access performance metrics, you need
--cap-add SYS_ADMIN - To access ngingx server (if you are running a demo), you need to forward
8080port, see-p 8080:8080
Summarizing, start container as follows (-v option maps a host folder to the container so you can copy transcoded streams back to the host):
DEVICE=${DEVICE:-/dev/dri/renderD128}
DEVICE_GRP=$(stat --format %g $DEVICE)
mkdir -p /tmp/media-delivery && chmod -R 777 /tmp/media-delivery
docker run --rm -it -v /tmp/media-delivery:/opt/media-delivery \
-e DEVICE=$DEVICE --device $DEVICE --group-add $DEVICE_GRP \
--cap-add SYS_ADMIN \
-p 8080:8080 \
intel-media-deliveryOnce inside a container you can run the included software and scripts. To start, we recommend running simple scripts which will showcase basic transcoding capabilities. These scripts will download sample video clips, though you can supply your own as a script argument if needed. If you work under proxy do not forget to add it to your environment (via export https_proxy=<...>).
Below commands will run single transcoding session (1080p or 4K) and produce output files which you can copy to the host and review:
# AV1 to AV1: ffmpeg-qsv-AV1-1080p.sh 1 ffmpeg-qsv-AV1-4K.sh 1 sample-multi-transcode-AV1-1080p.sh 1 sample-multi-transcode-AV1-4K.sh 1 # AVC to AVC: ffmpeg-qsv-AVC-1080p.sh 1 ffmpeg-qsv-AVC-4K.sh 1 sample-multi-transcode-AVC-1080p.sh 1 sample-multi-transcode-AVC-4K.sh 1 # HEVC to HEVC ffmpeg-qsv-HEVC-1080p.sh 1 ffmpeg-qsv-HEVC-4K.sh 1 sample-multi-transcode-HEVC-1080p.sh 1 sample-multi-transcode-HEVC-4K.sh 1Below commands will run specified number of parallel transcoding sessions (1080p or 4K). No output files will be produced, but you can check performance. Mind that below numbers of parallel transcoding sessions are suggested for Intel® Data Center GPU Flex Series. Other GPUs might support different number of sessions running at realtime:
# AV1 to AV1: ffmpeg-qsv-AV1-1080p.sh 16 ffmpeg-qsv-AV1-4K.sh 4 sample-multi-transcode-AV1-1080p.sh 16 sample-multi-transcode-AV1-4K.sh 4 # AVC to AVC: ffmpeg-qsv-AVC-1080p.sh 12 ffmpeg-qsv-AVC-4K.sh 2 sample-multi-transcode-AVC-1080p.sh 12 sample-multi-transcode-AVC-4K.sh 2 # HEVC to HEVC ffmpeg-qsv-HEVC-1080p.sh 16 ffmpeg-qsv-HEVC-4K.sh 4 sample-multi-transcode-HEVC-1080p.sh 16 sample-multi-transcode-HEVC-4K.sh 4
These scripts run transcoding command lines which we recommend to use for best performance and quality in case of Random Access encoding. See reference command lines for details.
In addition to the simple scripts described above, this project provides the following scripts and software which can be tried next:
For the more complex samples, check out Open Visual Cloud and their full scale CDN Transcode Sample.
Intel® Deep Link Hyper Encode Technology was designed to boost transcode performance to achieve 8K60 real-time throughput. Currently supported via Sample Multi Transcode application, this technology was tested on Intel® Data Center GPU Flex 140 card (2 GPU nodes). Media delivery container can be used to check two different flavors of this approach:
- 1 GPU node solution (encoder and decoder workloads are shared between 2 VDBOX engines of a single GPU node and encoding is parallelized on a GOP level)
- 2 GPU node solution (encoder and decoder workloads use 4 VDBOX engines of 2 GPU nodes and encoding is parallelized on a GOP level)
Simple example scripts are provided allowing for a quick test of 8k60 transcoding at the user’s end. First, start docker with mapping multiple GPU nodes (mind --device /dev/dri vs --device $DEVICE as in previous examples) as follows:
DEVICE=${DEVICE:-/dev/dri/renderD128}
DEVICE_GRP=$(stat --format %g $DEVICE)
mkdir -p /tmp/media-delivery && chmod -R 777 /tmp/media-delivery
docker run --rm -it -v /tmp/media-delivery:/opt/media-delivery \
-e DEVICE=$DEVICE --device /dev/dri --group-add $DEVICE_GRP \
--cap-add SYS_ADMIN \
-p 8080:8080 \
intel-media-delivery Once inside the container, simple scripts can be used to showcase Intel® Deep Link Hyper Encode Technology. If you work behind a firewall, please add HTTPS proxy to your environment (via export https_proxy=<...>) before running the scripts.
Use the following commands to run single 8K transcoding session using either of the two flavors:
# AV1 to AV1: sample-multi-transcode-AV1-8K-hyperenc1gpu.sh sample-multi-transcode-AV1-8K-hyperenc2gpu.sh # HEVC to HEVC sample-multi-transcode-HEVC-8K-hyperenc1gpu.sh sample-multi-transcode-HEVC-8K-hyperenc2gpu.sh
More information on Intel® Deep Link Hyper Encode Technology can be found here:
Feedback and contributions are welcome. Please, submit issues and pull requests here at GitHub.
Dockerfiles should be supported as described in the document.
If changes are done to dockerfiles and/or scipts, please, make sure to run tests before submitting pull requests.
Container image comes with some embedded content attributed as follows:
/opt/data/embedded/WAR_TRAILER_HiQ_10_withAudio.mp4:
Film: WAR - Courtesy & Copyright: Yash Raj Films Pvt. Ltd.Inside the container, please, refer to the following file:
cat /opt/data/embedded/usage.txt