This repository allows one to use TensorRT in Unity. Currently, it is at a pre-alpha level (i.e, it barely just works, and could be significantly optimized).
- Load an ONNX model and perform inference on it, using Unity Texture2D's as input and output. This means there's no GPU<->CPU copies, the only copy is from texture memory to linear memory and back (to run the model), which is only on the GPU.
- Serialize and deserialize TensorRT engines to speed up startup time.
- Allow multiple models to be used at once.
- Benchmark vs https://github.com/keijiro/Pix2Pix .
- Currently only supports DirectX 11 (easy to solve) and Windows 10 (harder to solve).
- Dispose might not clean up all memory.
- If you're using CUDA 9.0 and Windows 10, you can use the Unity project (the "tensor rt" folder) as is. It uses the
unitytrtcuda.dllincluded in the unitytrtcuda/x64/Release. - If not, you'll have to build from source.
- Install Visual Studio 2017 and Visual Studio v140 build tools (Re-run the installer, click on modify, and check the box for it if you already installed Visual Studio).
- Install CUDA. It should also install the Visual studio build tools for CUDA. You can check this worked by if you click
New Projectit gives you an option to use the CUDA template for a project. - Download TensorRT. It's free, but you have to sign-up. This code was built with TensorRT 5.03, not sure how it'll fare with earlier or later versions.
- Open the Visual Studio project in the unitytrtcuda folder. Modify the project's include directories and the linker's Additional Libraries to point at your downloaded TensorRT.
- Compile. This should produce a DLL. Copy this DLL, the TensorRT DLLs, as well as the copied CUDA DLLs in the output folder, to the Unity project's Assets/Plugins folder. Note that you have to restart unity for DLL changes to kick in.