Enable neural graphics using ML Extensions for Vulkan
Introduction
Run neural graphics workloads with ML Extensions for Vulkan
Setting up the ML Emulation Layers for Vulkan
Simple Tensor and Data Graph
Running a test with the Scenario Runner
Use RenderDoc to debug and analyze workloads
Wrapping up
Next Steps
Enable neural graphics using ML Extensions for Vulkan
Debug and profile workloads with RenderDoc
Integrating machine learning into real-time rendering makes frame-level inspection and performance analysis critical. RenderDoc helps you visualize and debug the workloads by letting you step through frames, examine tensors, and inspect Vulkan API calls.
RenderDoc is a powerful GPU frame capture tool that lets you:
- Step through a frame’s rendering process
- Inspect Vulkan API calls
- View shader inputs and outputs
- Examine GPU resource states and memory usage
When to use RenderDoc with the samples
RenderDoc can help in scenarios such as:
- Diagnosing unexpected visual output by stepping through draw calls
- Analyzing the order and behavior of Vulkan API calls
- Investigating memory consumption or GPU resource state
- Validating execution of data graph pipelines or identifying sync issues
Installing Arm Performance Studio (includes RenderDoc)
To use RenderDoc with ML extensions, install the Arm-customized version via Performance Studio:
- Download Arm Performance Studio from the
Arm Developer website
. The minimum version to use is
2025.4 - Run the installer:
Arm_Performance_Studio_<version>_windows_x86-64.exe - Follow the installation instructions.
Once installed, launch RenderDoc for Arm GPUs via the Windows Start menu.
Capture Vulkan frames with RenderDoc
You can capture and inspect Vulkan Samples that use ML extensions for Vulkan, including API calls such as vkCreateTensorARM and structures like VK_STRUCTURE_TYPE_TENSOR_DESCRIPTION_ARM.
RenderDoc is especially useful for visualizing tensor operations, inspecting resource bindings, and verifying correct data graph pipeline execution.
Capture with RenderDoc
- Open RenderDoc, and in the main window, go to the Launch Application section.
- Configure the following fields:
- Executable Path: Path to the built executable
vulkan_samples.exe. - Working Directory: Path to the root of the Vulkan Samples project.
- Command-line Arguments:
You can substitutesample simple_tensor_and_data_graphsimple_tensor_and_data_graphwith any of the other sample names as needed.
- Executable Path: Path to the built executable
- Click Launch. The selected sample will start running.
- Once the application window is active, press F12 to capture a frame.
- After the frame is captured, it will appear in RenderDoc’s capture list. Double-click it to explore the captured frame and inspect ML extensions for Vulkan calls in detail.
Learn more
This workflow enables close inspection of how ML graphs are built and executed within Vulkan — an essential tool when optimizing pipelines or debugging integration issues. If you want to learn more about RenderDoc for Arm GPUs, you can check out the Debug With RenderDoc User Guide .
Move on to the next section to review further resources on what is new, and what is coming.