This project includes Dockerfiles for building a Vulkan SDK container that can be used with NVIDIA GPUs.
Before getting started, ensure that the system has the latest NVIDIA driver and the NVIDIA Container Toolkit installed.
Clone the git repository and then build the image:
docker build --pull -t \
nvidia/vulkan:1.3-470 \
--build-arg BASE_DIST=ubuntu20.04 \
--build-arg CUDA_VERSION=11.4.2 \
--build-arg VULKAN_SDK_VERSION=`curl -sk https://vulkan.lunarg.com/sdk/latest/linux.txt` \
--file docker/Dockerfile.ubuntu .The Vulkan SDK container is typically intended to be used as the base-layer for other Vulkan applications.
To see an example, refer to the samples/ directory, which showcases how the image can be used in multi-stage
Docker builds.
The Vulkan SDK container however includes utilities such as vulkaninfo, which can be used to test whether
the GPUs can be enumerated by the Vulkan loader and driver.
As a first step, launch the container and open a terminal.
NOTE: Some of the loaders need to be currently specified on the command line with Docker as bind mounts. This issue will be fixed in a future release of the NVIDIA Container Toolkit.
A common utility to test Vulkan is to run vulkaninfo. In this example, we run vulkaninfo without
specifying any DISPLAY variables:
docker run --gpus all \
-e NVIDIA_DISABLE_REQUIRE=1 \
-e NVIDIA_DRIVER_CAPABILITIES=all --device /dev/dri \
-v /etc/vulkan/icd.d/nvidia_icd.json:/etc/vulkan/icd.d/nvidia_icd.json \
-v /etc/vulkan/implicit_layer.d/nvidia_layers.json:/etc/vulkan/implicit_layer.d/nvidia_layers.json \
-v /usr/share/glvnd/egl_vendor.d/10_nvidia.json:/usr/share/glvnd/egl_vendor.d/10_nvidia.json \
-it nvidia/vulkan:1.3-470 \
bashOnce inside the container, run vulkaninfo:
root@075caf680108:/# vulkaninfoYou should see an output such as below, which shows the version, enumerated GPUs and supported NVIDIA extensions.
==========
VULKANINFO
==========
Vulkan Instance Version: 1.3.204
...
VK_LAYER_NV_optimus (NVIDIA Optimus layer) Vulkan version 1.2.175, layer version 1:
Layer Extensions: count = 0
Devices: count = 1
GPU id = 0 (Tesla T4)
Layer-Device Extensions: count = 0
...
Device Properties and Extensions:
=================================
GPU0:
VkPhysicalDeviceProperties:
---------------------------
apiVersion = 4202671 (1.2.175)
driverVersion = 1972256896 (0x758e4080)
vendorID = 0x10de
deviceID = 0x1eb8
deviceType = PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
deviceName = Tesla T4
...
VK_NV_inherited_viewport_scissor : extension revision 1
VK_NV_mesh_shader : extension revision 1
VK_NV_ray_tracing : extension revision 3
VK_NV_representative_fragment_test : extension revision 2
VK_NV_sample_mask_override_coverage : extension revision 1
VK_NV_scissor_exclusive : extension revision 1
VK_NV_shader_image_footprint : extension revision 2
VK_NV_shader_sm_builtins : extension revision 1
VK_NV_shader_subgroup_partitioned : extension revision 1
VK_NV_shading_rate_image : extension revision 3
VK_NV_viewport_array2 : extension revision 1
VK_NV_viewport_swizzle : extension revision 1
Another graphical utility to test the Vulkan setup is to run the vkcube application. This utility spins
up a window on the client side, so a pre-requisite is to ensure that there is an X server running on the client
side.
First, allow users to access the X server:
$ xhost +Then, specify the DISPLAY environment variable before starting the container:
docker run --gpus all \
-e NVIDIA_DISABLE_REQUIRE=1 \
-v $HOME/.Xauthority:/root/.Xauthority \
-e DISPLAY -e NVIDIA_DRIVER_CAPABILITIES=all --device /dev/dri --net host \
-v /etc/vulkan/icd.d/nvidia_icd.json:/etc/vulkan/icd.d/nvidia_icd.json \
-v /etc/vulkan/implicit_layer.d/nvidia_layers.json:/etc/vulkan/implicit_layer.d/nvidia_layers.json \
-v /usr/share/glvnd/egl_vendor.d/10_nvidia.json:/usr/share/glvnd/egl_vendor.d/10_nvidia.json \
-it nvidia/vulkan:1.3-470 \
bashOnce inside the container, run the vkcube utility:
root@075caf680108:/# vkcube
Selected GPU 0: Tesla T4, type: 2And then on the client side, you should be able to observe a spinning cube in a window:
A sample container is provided to test headless Vulkan functionality - this is especially useful
on NVIDIA GPUs without display (i.e. datacenter class products). Two samples are included - computeheadless,
which computes Fibonacci series and a renderheadless, which renders an image and saves it to local storage.
A sample container has already been pre-built to test the Vulkan functionality. Simply run as:
docker run --gpus all \
dualvtable/vulkan-sample:latestThe container will run through the headless samples with an output shown below:
Running headless compute example
GPU: Tesla T4
Compute input:
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Compute output:
0 1 1 2 3 5 8 13 21 34 55 89 144 233 377 610 987 1597 2584 4181 6765 10946 17711 28657 46368 75025 121393 196418 317811 514229 832040 1346269
Finished. Press enter to terminate...\n
Running headless rendering example
GPU: Tesla T4
Framebuffer image saved to headless.ppm
Finished. Press enter to terminate...\nThe sample may also be built by cloning the Dockerfile available in the samples/ directory
of the repository.
cd samples/ \
&& docker build --pull -t \
nvidia/vulkan-sample \
--build-arg VULKAN_SDK_VERSION=`curl -sk https://vulkan.lunarg.com/sdk/latest/linux.txt | awk '{ printf("%.2g", $1) }` \
--build-arg DRIVER_VERSION=470 \
--file Dockerfile.ubuntu .The Vulkan SDK image is versioned as <Vulkan-SDK>-<Driver-Branch>.
The source code in this repository is licensed under Apache 2.0. The resulting container images are licensed under the NGC Deep Learning Container License.
Checkout the Contributing document!
- Please let us know by filing a new issue
- You can contribute by opening a pull request on the gitlab repository