rolaand-jayz/re-of-fsr-4.1.0-upscaling

AMD FidelityFX Super Resolution 4.1.0, known as FSR 4.1.0, is a technology built into modern graphics cards that uses a small neural network to reconstruct high-resolution images from lower-resolution inputs. AMD ships this technology as compiled Windows DLL files with no public documentation of how the internal pipeline works. This repository documents a static analysis of those files, meaning the author examined the raw binary code without ever running it, to understand the structure of the neural network inside. The analysis found and cataloged 602 shader programs (small GPU programs that process pixel data) embedded in the DLL, and identified 27 distinct passes the neural network runs through when upscaling an image. It also extracted 6 weight blobs, which are the actual learned values that make the neural network work, each 131,072 bytes in size. The format of those weight blobs was decoded, revealing 16-bit bias values, 8-bit weight values, and a small region of 32-bit output values. The tools used include Ghidra (a reverse-engineering tool released by the NSA), DXIL (the intermediate shader language for DirectX), and raw x86-64 disassembly. The repository is careful to mark which findings are confirmed from static analysis and which are only inferred. Things the project can prove include the shader catalog, the weight blob contents, and the basic dispatch loop structure. Things it cannot prove include the actual order the passes run in during a real game session, the exact values fed to each pass, and whether any reconstructed output would match AMD's original binary behavior. A dedicated validation document tracks the status of every claim. The audience is researchers who want to understand how FSR 4.1.0 works internally, developers studying GPU pipeline architecture, or anyone interested in neural-network-based image upscaling. The project is not a replacement for AMD's driver and is not meant to be run in games. The full README is longer than what was shown.