This sample illustrates DirectML and ONNX Runtime inference of the ESRGAN super-resolution model (converted to ONNX by Qualcomm AI Hub). While the model is converted from PyTorch using an IHV-specific toolchain, the same ONNX model can run on either GPU or NPU devices across hardware vendors when using DirectML. You may wish to use this sample as a simple template for learning:

• How to build a C++ application that depends a prebuilt release of ONNX Runtime with the DirectML execution provider (EP).
• How to use ONNX Runtime's C++ API to run inference of ONNX models with the DirectML EP. This sample illustrates the simpler path of binding CPU-side tensor data, which is automatically converted to DirectX resources within ONNX Runtime (directly binding of D3D resources is not shown).
• How to enumerate and select a specific DirectX adapter (GPU or NPU) to be used by the DirectML EP.
• How to load and convert image files to NCHW tensors compatible with ONNX Runtime by leveraging WIC, an API built into Windows.

The included variant of the ESRGAN model accepts a 128x128 input image and upscales it by factor of 4X to 512x512. The images below compare the low-resolution input to its high-resolution output.

[Original Image] An 900x659 test image (Grevy's Zebra Stalion by Rainbirder, CC BY-SA 2.0 https://creativecommons.org/licenses/by-sa/2.0, via Wikimedia Commons).

[Model Input] The test image cropped and downsampled to 128x128 using a high-quality bicubic filter. When viewing this image in this README it will be upscaled to 512x512 for easier comparison. This image is converted to a NCHW tensor and passed into the model for inference.

[Model Output] The output NCHW float32 tensor converted back to an image, which reconstructs significantly more detail than a simple bilinear or bicubic filter.

This sample uses CMake to generate a Visual Studio solution and download external dependencies. The following instructions assume you have installed CMake and it is available on the PATH.

Run CMake to configure and build the sample:

> cmake --preset win-<arch>
> cmake --build --preset win-<arch>-<config>

• <arch> must be one of: x64, arm64.
• <config> must be one of: <release>, <debug>.

Running the compiled executable will list all enumerate DX adapters, select a DX adapter (by default the first), then use this adapter to run inference with the model:

> cd build/win-<arch>/<config>
> .\directml_esrgan.exe

Adapter[0]: NVIDIA GeForce RTX 4080 (SELECTED)
Adapter[1]: Intel(R) UHD Graphics 770
Adapter[2]: Microsoft Basic Render Driver
Saving cropped/scaled image to input.png
Saving inference results to output.png

You may use the -a <adapter_substring> to select a different adapter that contains part of the <adapter_substring> in its description. For example, you can run on a compatible¹ NPU using -a NPU:

> cd build/win-arm64/release
> .\directml_esrgan.exe -a NPU
Adapter[0]: Snapdragon(R) X Elite - X1E78100 - Qualcomm(R) Adreno(TM) GPU
Adapter[1]: Snapdragon(R) X Elite - X1E78100 - Qualcomm(R) Hexagon(TM) NPU (SELECTED)
...

You must have Windows 11 24H2 installed to run this sample using an NPU.

¹ Currently, this sample is only supported with the NPU available in Copilot+ PCs with Snapdragon X Series processors. Support for additional NPUs will come soon. To run this sample with the QC NPU, refer to our blog post for list of technical requirements required.

² The following messages "Error in cpuinfo" and "Unknown Qualcomm CPU" are mistaken messages which will be fixed in a future release. They are expected, and the model has executed successfully.

Error in cpuinfo: Unknown chip model name 'snapdragon (tm) 8cx gen 3 @ 3.40 GHz'.

unknown Qualcomm CPU part 0x1 ignored