FidelityFX Super Resolution 1.0: Technical Analysis and Implementation

This article provides a comprehensive technical analysis of AMD's FidelityFX Super Resolution (FSR) 1.0, a spatial upscaling algorithm that consists of two main components: EASU (Edge Adaptive Spatial Upsampling) and RCAS (Robust Contrast Adaptive Sharpening). The article begins by explaining the Lanczos2 sampling function and its polynomial fitting approach used in FSR 1.0 to optimize performance by avoiding expensive trigonometric calculations.

The EASU algorithm workflow is detailed, including sampling 12 surrounding points, calculating luminance, performing four directional calculations using bilinear interpolation to obtain gradients and lengths, calibrating these values, rotating based on direction, and computing Lanczos2 weights. The article explains the 12-point sampling pattern and how the algorithm processes four '+'-shaped patterns to calculate direction vectors and lengths.

The numerical calibration section describes the mapping relationship between features and the Lanczos fitting polynomial control factor ω. The code flow is illustrated with detailed function calls and mathematical formulas for rotation and weighted averaging. The article explains how EASU handles both edge and non-edge pixels using a unified formula, with non-edge pixels using weighted averaging and edge pixels using high-frequency filtering with Lanczos2.

The RCAS component is explained as a sharpening step that uses a Laplacian operator variant, with the weight w calculated based on surrounding pixel contrast. The article concludes with optimization techniques for mobile deployment, including using half-precision, reducing texture sampling, early exit strategies, removing deringing, and optimizing the FsrEasuSetF function calls.

Overall, this is a detailed technical breakdown of FSR 1.0's implementation, providing valuable insights for graphics programmers and researchers working on upscaling algorithms and real-time rendering optimization.