The Simple Tensor and Data Graph sample is your starting point for working with the ML extensions for Vulkan. It demonstrates how to execute a simple neural network using a data graph pipeline — specifically, a 2D average pooling operation.
With the environment set up, you can now grab the sample code. These examples are maintained in a fork of the Khronos Group’s repository.
git clone --recurse-submodules https://github.com/ARM-software/Vulkan-Samples --branch tensor_and_data_graph
cd Vulkan-Samples
This repository includes the framework and samples showcasing the ML extensions for Vulkan.
You’re now ready to compile the project. From the root of the repository:
Be sure to run the commands in Developer Mode to avoid permission issues.
Generate Visual Studio project files using CMake:
cmake -G "Visual Studio 17 2022" -A x64 -S . -Bbuild/windows
Finally, compile the vulkan_samples target in Release mode:
cmake --build build/windows --config Release --target vulkan_samples
Run the built executable using the following command:
build\windows\app\bin\Release\AMD64\vulkan_samples.exe sample simple_tensor_and_data_graph
This should open a new window visualizing the operation. In this sample, a minimal Vulkan application sets up a data graph pipeline configured to process a small neural network.
The sample creates input and output tensors, binds them using descriptor sets and pipeline layouts, and supplies a SPIR-V module that defines the network operation. Finally, it records and dispatches commands to execute the pipeline — and visualizes the results in real time. More details about what’s going on under the hood can be found in the documentation .
By running this sample, you’ve stepped through a complete Vulkan data graph pipeline powered by the ML extensions for Vulkan. You’ve created tensors, set up descriptors, built a SPIR-V-encoded ML graph, and dispatched inference — all without needing custom shaders. This sets the foundation for neural graphics. As you explore the remaining samples, you’ll see how this core pattern extends into real-world graphics scenarios.
As a next step, you can explore the remaining samples for the data graph pipeline. The documentation sits in each sample’s directory, available under samples/extensions/tensor_and_data_graph/ in the repository.
| Sample Name | Description | Focus Area |
|---|---|---|
| Graph Constants | Shows how to include constants like weights and biases into the data graph pipeline using tensors | Constant tensor injection |
| Compute Shaders with Tensors | Demonstrates how to feed tensor data into or out of data graph pipelines using compute shaders | Shader interoperability |
| Tensor Image Aliasing | Demonstrates tensor aliasing with Vulkan images to enable zero-copy workflows | Memory-efficient data sharing |
| Postprocessing with VGF | Explores using VGF format, which contains SPIR-V, input, output and constant data used to run a data graph pipeline. | Neural network model |
Next, you’ll review additional tools to help you work with ML extensions for Vulkan in your own development environment.