This repository contains the source code for a sample application that demonstrates the use of Cooperative Vectors, a new feature introduced in DirectX 12. This feature enables standardized, cross-vendor access to hardware-accelerated matrix multiplication units.

The demo was presented at the Game Developers Conference (GDC) 2025, as part of the Advanced Graphics Summit sessions.

You can find some additional informations here.

• A GPU that supports DirectX 12 Cooperative Vectors, such as:
  • Intel® Graphics LunarLake (or newer)
  • Intel® Arc B-Series Graphics Cards (or newer)
• A driver that supports DirectX 12 Cooperative Vectors (or newer)
• Enable Windows Developer Mode
• Compiler with C++20 support
  • Microsoft Visual Studio 2022 or newer
• CMake 3.8 or newer

git clone https://github.com/GameTechDev/TextureSetNeuralCompressionSample.git
cd TextureSetNeuralCompressionSample

Download the Microsoft® DirectX 12 Agility SDK version 717 (preview or later) and the Microsoft® DirectX Compiler version 8.2502.8 (or later). You can do this manually or run the dependencies.bat script to automatically download and install the required dependencies.

mkdir build
cd build
cmake .. -DDX12_SDK_VERSION=717
cmake --build . --config Release

The sample executable is called dino_danger. You can run it from the Microsoft Visual Studio 2022 IDE or via the command line:

cd output/bin/Release
dino_danger.exe --data-dir ../../..

You can also get more info about the command line options by running:

dino_danger.exe --help
    --data-dir Location of the resource folders.
    --adapter-id Integer that allows to pick the desired GPU [-1 = Largest VRAM, >= 0 System adapter ID].
    --poi Integer that allows to pick the initial camera location.
    --disable-coop Disable cooperative vector usage at launch.
    --disable-animation Disable mesh animation at launch.
    --rendering-mode Pick the rendering mode [0 = Material, 1 = GBuffer, 2 = Debug].
    --texture-mode Pick the texture mode [0 = Uncompressed, 1 = BC6, 2 = Neural].
    --filtering-mode Pick the filtering mode [0 = Nearest, 1 = Linear, 2 = Anisotropic].

Intel® Texture Set Neural Compression Sample is licensed under the MIT License.