Neural Hijacking: Decoding DLSS5-Autopilot's Runtime Injection Strategy for Legacy Game Support

Title: Neural Hijacking: Decoding DLSS5-Autopilot's Runtime Injection Strategy for Legacy Game Support

Original Source: <source 3adb506d-66f>

Publication Date: Originally published on tamiz.pro

Summary:

DLSS5-Autopilot, a hypothetical advanced modular framework, aims to bring Deep Learning Super Sampling (DLSS) to legacy games that lack native support. This article analyzes the runtime injection strategy employed by DLSS5-Autopilot to achieve this feat, focusing on its architectural components and dynamic shader hot-swapping mechanism.

Architectural Overview:

DLSS5-Autopilot operates at the OS driver boundary, bypassing the need for game engine modification. It consists of three layers:

1. Interception Layer: A kernel-mode driver extension or user-space hooking library that intercepts DirectX 12/Vulkan calls, such as CreateDevice, CreateCommandQueue, and EndFence.

2. Inference Engine: A background process that manages the neural network models for upscaling, sharpening, and motion vector estimation.

3. Shader Hot-Swap Module: A runtime assembler that patches the game's post-processing shaders to route the final frame through the DLSS inference pipeline.

Dynamic Shader Hot-Swapping Mechanism:

The core challenge in backporting DLSS to unsupported games is the lack of motion vectors. DLSS5-Autopilot addresses this with a Dual-Stage Optical Flow Estimator.

1. Shader Graph Analysis: The hooking library captures root signatures and shader bytecode during game initialization. It then identifies and parses the Final Resolve pass in the Pipeline State Object (PSO).

2. Runtime Assembly Injection: Instead of replacing the shader entirely, the system uses tail-call injection. It appends a new block of machine code to the existing shader's entry point, which signals the inference engine, waits for the DLSS model to complete, and samples the upscaled texture to output the result.

Motion Vector Estimation:

DLSS requires motion vectors for temporal accumulation, but legacy games often lack these vectors. DLSS5-Autopilot uses a Dual-Stage Optical Flow Estimator to estimate motion vectors from the previous frame, enabling DLSS backporting.

Key Points:

- DLSS5-Autopilot enables DLSS on legacy games by operating at the OS driver level, bypassing the need for game engine integration.

- It consists of an Interception Layer, Inference Engine, and Shader Hot-Swap Module.

- The Shader Hot-Swap Module uses tail-call injection to replace the game's upscaling shader without crashing the application.

- DLSS5-Autopilot employs a Dual-Stage Optical Flow Estimator to estimate motion vectors in unsupported games, enabling temporal accumulation for DLSS.