OpenCL Acceleration for Azure Docker Containers

Much of my work relies on accelerated computation of some sort, usually applied to images or other 2D/3D data. While CUDA is the de-facto toolkit in my field, OpenCL is a more intriguing solution thanks to its portability. This is additionally important in a cloud environment where the ability to run the same code/containers one a variety of different node SKUs is very appealing.

This post walks through my experiments with bringing multi-platform OpenCL support to Azure Kubernetes via Docker.

There are tons of existing OpenCL dockerfiles demonstrating multiplatform support. (ex: here for the one I used as a base, and plenty more). I wanted:

1. Based on Ubuntu 20.04LTS
2. Support for at least Intel and CUDA OpenCL (with AMD GPUs as a bonus)
3. Set up for use as a VSCode DevContainer Image

The result was a Dockerfile that looks like the following:

FROM nvidia/cuda:11.6.0-devel-ubuntu20.04
ARG DEBIAN_FRONTEND=noninteractive

################################ begin nvidia opencl driver ################################

RUN apt-get update && apt-get install -y --no-install-recommends \
        ocl-icd-libopencl1 \
        clinfo pkg-config && \
    rm -rf /var/lib/apt/lists/*

RUN mkdir -p /etc/OpenCL/vendors && \
    echo "libnvidia-opencl.so.1" > /etc/OpenCL/vendors/nvidia.icd

RUN echo "/usr/local/nvidia/lib" >> /etc/ld.so.conf.d/nvidia.conf && \
    echo "/usr/local/nvidia/lib64" >> /etc/ld.so.conf.d/nvidia.conf

ENV PATH /usr/local/nvidia/bin:${PATH}
ENV LD_LIBRARY_PATH /usr/local/nvidia/lib:/usr/local/nvidia/lib64:${LD_LIBRARY_PATH}

# nvidia-container-runtime
ENV NVIDIA_VISIBLE_DEVICES all
ENV NVIDIA_DRIVER_CAPABILITIES compute,utility

################################ end nvidia opencl driver ################################

############### begin Intel OpenCL driver installation ###############
RUN apt-get update && apt-get install -y alien clinfo pkg-config

# Install Intel OpenCL driver
#ENV INTEL_OPENCL_URL=http://registrationcenter-download.intel.com/akdlm/irc_nas/vcp/13793/l_opencl_p_18.1.0.013.tgz
ENV INTEL_OPENCL_URL=http://registrationcenter-download.intel.com/akdlm/irc_nas/9019/opencl_runtime_16.1.1_x64_ubuntu_6.4.0.25.tgz

RUN mkdir -p /tmp/opencl-driver-intel
WORKDIR /tmp/opencl-driver-intel
RUN curl -O $INTEL_OPENCL_URL; \
    tar -xzf $(basename $INTEL_OPENCL_URL); \
    for i in $(basename $INTEL_OPENCL_URL .tgz)/rpm/*.rpm; do alien --to-deb $i; done; \
    dpkg -i *.deb; \
    mkdir -p /etc/OpenCL/vendors; \
    echo /opt/intel/*/lib64/libintelocl.so > /etc/OpenCL/vendors/intel.icd; \
    rm -rf *

############### end Intel OpenCL driver installation ###############

# Update & install packages
RUN apt-get update && \
    apt-get install -y wget make clinfo build-essential git libcurl4-openssl-dev libssl-dev zlib1g-dev libcurl4-openssl-dev libssl-dev

To use this in VSCode, a .devcontainer folder is first created in the root of the project you’re going to be working with (for me, a new empty git repository). Inside the .devcontainer folder, you need to put the above Dockerfile as well as a devcontainer.json configuration file that tells VSCode how to configure your development environment. Mine looks like the following:

{
	"name": "${localWorkspaceFolderBasename}",
	"dockerFile": "Dockerfile",

	// Set *default* container specific settings.json values on container create.
	"settings": { 
		"terminal.integrated.shell.linux": "/bin/bash"
	},

	// To enable your local GPUs in container if they are on enabled by default
	"runArgs": [ "--gpus=all" ],

	// Use 'forwardPorts' to make a list of ports inside the container available locally.
	"forwardPorts": [9002],

	"extensions": [
		"ms-vscode.cpptools",
		"ms-vscode.cmake-tools",
		"eamodio.gitlens",
	],

	// Create and mount the working directory in the container
	//"postCreateCommand": "apt-get update && apt-get -y install gdb"
}

Once both of these files are present and saved, you can hit “F1”, then select “Remote-Containers: Rebuild Container Without Cache” (you’ll need to have the Remote Container extension and Docker installed). Once the DevContainer is built, you’ll be greeted with a new empty terminal in VSCode, and will be ready to get started developing.