Optimizing Vulkan Dispatch Schedules for Real-Time U-Net Denoising

Image denoising is fundamental to Monte Carlo rendering. Recently, real-time path-traced applications have become viable, relying heavily on efficient denoising. Modern denoisers are often based on neural networks, most commonly variants of the U-Net architecture. While tooling for development and training of custom neural networks is well established in Python, existing deployment and interop strategies typically rely on external inference frameworks, incur host synchronization, data movement, or heavyweight dependencies, making them impractical for tightly integrated, real-time rendering pipelines. We present a light-weight and portable ONNX to Vulkan conversion framework that is designed for efficient, fully GPU-resident deployment of U-Net-based networks. We employ graph-based and profile-guided optimizations to determine optimal Vulkan dispatch schedules, including block sizes, data layouts, and kernel fusion decisions. Our method includes effective pruning of the vast search space, significantly reducing compile times. Using the Open Image Denoise (OIDN) network, we demonstrate runtime improvements over established inference frameworks such as TensorRT.