Media Reference Stack¶
The Media Reference Stack (MeRS) is a highly optimized software stack for Intel® Architecture Processors (the CPU) and Intel® Processor Graphics (the GPU) to enable media prioritized workloads, such as transcoding and analytics.
This guide explains how to use the pre-built MeRS container image, build your own MeRS container image, and use the reference stack.
Developers face challenges due to the complexity of software integration for media tasks that require investing time and engineering effort. For example:
Finding the balance between quality and performance.
Understanding available standard-compliant encoders.
Optimizing across the hardware-software stack for efficiency.
MeRS abstracts away the complexity of integrating multiple software components and specifically tunes them for Intel platforms. MeRS enables media and visual cloud developers to deliver experiences using a simple containerized solution.
Refer to the System Stacks for Linux* OS repository for information and download links for the different versions and offerings of the stack.
MeRS V0.2.0 release announcement including media processing on GPU and analytics on CPU.
MeRS V0.1.0 including media processing and analytics CPU.
MeRS Release notes on Github* for the latest release of Deep Learning Reference Stack
MeRS can run on any host system that supports Docker*. This guide uses Clear Linux* OS as the host system.
To install Clear Linux OS on a host system, see how to install Clear Linux* OS from the live desktop.
To install Docker* on a Clear Linux OS host system, see the instructions for installing Docker*.
For optimal media analytics performance, a processor with Vector Neural Network Instructions (VNNI) should be used. VNNI is an extension of Intel® Advanced Vector Extensions 512 (Intel® AVX-512) and is available starting with the 2nd generation of Intel® Xeon® Scalable processors, providing AI inference acceleration.
The MeRS provides a pre-built Docker image available on DockerHub, which includes instructions on building the image from source. MeRS is open-sourced to make sure developers have easy access to the source code and are able to customize it. MeRS is built using the latest clearlinux/os-core Docker image and aims to support the latest Clear Linux OS version.
MeRS provides the following libraries and drivers:
Scalable Video Technology for HEVC encoding, also known as H.265
Scalable Video Technology for AV1 encoding
x264 for H.264/MPEG-4 AVC encoding
dav1d for AV1 decoding
VAAPI (Video Acceleration API) open-source library (LibVA), which provides access to graphics hardware acceleration capabilities.
Intel® Media Driver for VAAPI for supporting hardware acceleration on Intel® Gen graphics hardware platforms.
Intel® Graphics Memory Management Library provides device specific and buffer management for the Intel® Graphics Compute Runtime for oneAPI Level Zero and OpenCL™ Driver and the Intel Media Driver for VAAPI.
Components of the MeRS include:
Clear Linux OS as a base for performance and security.
OpenVINO™ toolkit for inference.
FFmpeg* with plugins for:
GStreamer* with plugins for:
The MeRS is validated on 11th generation Intel Processor Graphics and newer. Older generations should work but are not tested against.
The pre-built MeRS container image configures FFmpeg without certain elements (specific encoder, decoder, muxer, etc.) that you may require. If you require changes to FFmpeg we suggest starting at Build the MeRS container image from source.
Pre-built MeRS Docker images are available on DockerHub* at https://hub.docker.com/r/sysstacks/mers-clearlinux
To use the MeRS:
Pull the image directly from Docker Hub.
docker pull sysstacks/mers-clearlinux
Once you have downloaded the image, run it using the following command:
docker run -it sysstacks/mers-clearlinux
This will launch the image and drop you into a bash shell inside the container. GStreamer and FFmpeg programs are installed in the container image and accessible in the default $PATH. Use these programs as you would outside of MeRS.
Paths to media files and video devices, such as cameras, can be shared from the host to the container with the --volume switch using Docker volumes.
If you choose to build your own MeRS container image, you can optionally add
customizations as needed. The
Dockerfile for the MeRS is available on
GitHub and can be used
as a reference when creating your own container image.
The MeRS image is part of the dockerfiles repository inside the Clear Linux OS organization on GitHub. Clone the
git clone https://github.com/intel/stacks.git
Navigate to the
stacks/mers/clearlinuxdirectory which contains the Dockerfile for the MeRS.
Use the docker build command with the
Dockerfileto build the MeRS container image.
docker build --no-cache -t sysstacks/mers-clearlinux .
This section shows examples of how the MeRS container image can be used to process media files.
The models and video source can be substituted from your use-case. Some publicly licensed sample videos are available at sample-videos repository for testing.
The examples below show transcoding using the GPU or CPU for processing.
On the host system, setup a workspace for data and models:
mkdir ~/ffmpeg mkdir ~/ffmpeg/input mkdir ~/ffmpeg/output
Copy a video file to
cp </path/to/video> ~/ffmpeg/input
Run the sysstacks/mers-clearlinux Docker image, allowing shared access to the workspace on the host:
docker run -it \ --volume ~/ffmpeg:/home/mers-user:ro \ --device=/dev/dri \ --env QSV_DEVICE=/dev/dri/renderD128 \ sysstacks/mers-clearlinux:latest
The --device parameter and the GSV_DEVICE environment variable allow shared access to the GPU on the host system. The values needed may be different depending on host’s graphics configuration.
After running the docker run command, you enter a bash shell inside the container.
From the container shell, you can run FFmpeg and GStreamer commands against the videos in
/home/mers-user/inputas you would normally outside of MeRS.
Some sample commands are provided for reference.
For more information on using the FFmpeg commands, refer to the FFmpeg documentation.
For more information on using the GStreamer commands, refer to the GStreamer documentation.
Example: Transcoding using GPU¶
The examples below show transcoding using the GPU for processing.
Using a FFmpeg to transcode raw content to SVT-HEVC and mp4:
ffmpeg -y -vaapi_device /dev/dri/renderD128 -f rawvideo -video_size 320x240 -r 30 -i </home/mers-user/input/test.yuv> -vf 'format=nv12, hwupload' -c:v h264_vaapi -y </home/mers-user/output/test.mp4>
Using a GStreamer to transcode H264 to H265:
gst-launch-1.0 filesrc location=</home/mers-user/input/test.264> ! h264parse ! vaapih264dec ! vaapih265enc rate-control=cbr bitrate=5000 ! video/x-h265,profile=main ! h265parse ! filesink location=</home/mers-user/output/test.265>
Example: Transcoding using CPU¶
The example below shows transcoding of raw yuv420 content to SVT-HEVC and mp4, using the CPU for processing.
ffmpeg -f rawvideo -vcodec rawvideo -s 320x240 -r 30 -pix_fmt yuv420p -i </home/mers-user/input/test.yuv> -c:v libsvt_hevc -y </home/mers-user/output/test.mp4>
Additional generic examples of FFmpeg commands can be found in the OpenVisualCloud repository and used for reference with MeRS.
This example shows how to perform analytics and inferences with GStreamer using the CPU for processing.
The steps here are referenced from the gst-video-analytics Getting Started Guide except simply substituting the gst-video-analytics docker image for the sysstacks/mers-clearlinux image.
The example below shows how to use the MeRS container image to perform video with object detection and attributes recognition of a video using GStreamer using pre-trained models and sample video files.
On the host system, setup a workspace for data and models:
mkdir ~/gva mkdir ~/gva/data mkdir ~/gva/data/models mkdir ~/gva/data/models/intel mkdir ~/gva/data/models/common mkdir ~/gva/data/video
Clone the opencv/gst-video-analytics repository into the workspace:
git clone https://github.com/opencv/gst-video-analytics ~/gva/gst-video-analytics cd ~/gva/gst-video-analytics git submodule init git submodule update
Clone the Open Model Zoo repository into the workspace:
git clone https://github.com/opencv/open_model_zoo.git ~/gva/open_model_zoo
Use the Model Downloader tool of Open Model Zoo to download ready to use pre-trained models in IR format.
If you are on a network with outbound proxies, you will need to configure set environment variables with the proxy server. Refer to the documentation on Proxy Configuration for detailed steps.
On Clear Linux OS systems you will need the python-extras bundle. Use sudo swupd bundle-add python-extras for the downloader script to work.
cd ~/gva/open_model_zoo/tools/downloader python3 downloader.py --list ~/gva/gst-video-analytics/samples/model_downloader_configs/intel_models_for_samples.LST -o ~/gva/data/models/intel
Copy a video file in h264 or mp4 format to
~/gva/data/video. Any video with cars, pedestrians, human bodies, and/or human faces can be used.
git clone https://github.com/intel-iot-devkit/sample-videos.git ~/gva/data/video
This example simply clones all the video files from the sample-videos repsoitory.
From a desktop terminal, allow local access to the X host display.
xhost local:root export DATA_PATH=~/gva/data export GVA_PATH=~/gva/gst-video-analytics export MODELS_PATH=~/gva/data/models export INTEL_MODELS_PATH=~/gva/data/models/intel export VIDEO_EXAMPLES_PATH=~/gva/data/video
Run the sysstacks/mers-clearlinux docker image, allowing shared access to the X server and workspace on the host:
docker run -it --runtime=runc --net=host \ -v ~/.Xauthority:/root/.Xauthority \ -v /tmp/.X11-unix:/tmp/.X11-unix \ -e DISPLAY=$DISPLAY \ -e HTTP_PROXY=$HTTP_PROXY \ -e HTTPS_PROXY=$HTTPS_PROXY \ -e http_proxy=$http_proxy \ -e https_proxy=$https_proxy \ -v $GVA_PATH:/home/mers-user/gst-video-analytics \ -v $INTEL_MODELS_PATH:/home/mers-user/intel_models \ -v $MODELS_PATH:/home/mers-user/models \ -v $VIDEO_EXAMPLES_PATH:/home/mers-user/video-examples \ -e MODELS_PATH=/home/mers-user/intel_models:/home/mers-user/models \ -e VIDEO_EXAMPLES_DIR=/home/mers-user/video-examples \ sysstacks/mers-clearlinux:latest
In the docker run command above:
--runtime=runc specifies the container runtime to be runc for this container. It is needed for correct interaction with X server.
--net=host provides host network access to the container. It is needed for correct interaction with X server.
/tmp/.X11-unixmapped to the container are needed to ensure smooth authentication with X server.
-v instances are needed to map host system directories inside the Docker container.
-e instances set the Docker container environment variables. Some examples need these variables set correctly in order to operate correctly. Proxy variables are needed if host is behind a firewall.
After running the docker run command, it will drop you into a bash shell inside the container.
From the container shell, run a sample analytics program in
~/gva/gst-video-analytics/samplesagainst your video source.
Below are sample analytics that can be run against the sample videos. Choose one to run:
Samples with face detection and classification:
./gst-video-analytics/samples/shell/face_detection_and_classification.sh $VIDEO_EXAMPLES_DIR/face-demographics-walking-and-pause.mp4 ./gst-video-analytics/samples/shell/face_detection_and_classification.sh $VIDEO_EXAMPLES_DIR/face-demographics-walking.mp4 ./gst-video-analytics/samples/shell/face_detection_and_classification.sh $VIDEO_EXAMPLES_DIR/head-pose-face-detection-female-and-male.mp4 ./gst-video-analytics/samples/shell/face_detection_and_classification.sh $VIDEO_EXAMPLES_DIR/head-pose-face-detection-male.mp4 ./gst-video-analytics/samples/shell/face_detection_and_classification.sh $VIDEO_EXAMPLES_DIR/head-pose-face-detection-female.mp4
When running, a video with object detection and attributes recognition (bounding boxes around faces with recognized attributes) should be played.
Sample with vehicle detection:
When running, a video with object detection and attributes recognition (bounding boxes around vehicles with recognized attributes) should be played.
Sample with FPS measurement:
The current version of MeRS does not include the Alliance for Open Media Video Codec (AOM). AOM can be built from source on an individual basis.
To add AOM support to the MeRS image:
The following programs are needed to add AOM support to MeRS: docker, git, patch. On Clear Linux OS these can be installed with the commands below. For other operating systems, install the appropriate packages.
sudo swupd bundle-add containers-basic dev-utils
Clone the Intel Stacks repository from GitHub.
git clone https://github.com/intel/stacks.git
Navigate to the directory for the MeRS image.
Apply the patch to the
patch -p1 < aom-patches/stacks-mers-v2-include-aom.diff
Use the docker build command to build a local copy of the MeRS container image tagged as aom.
docker build --no-cache -t sysstacks/mers-clearlinux:aom .
Once the build has completed successfully, the local image can be used following the same steps in this tutorial by substituting the image name with sysstacks/mers-clearlinux:aom.
Intel, Xeon, OpenVINO, and the Intel logo are trademarks of Intel Corporation or its subsidiaries. OpenCL and the OpenCL logo are trademarks of Apple Inc. used by permission by Khronos.