Unfortunately, information on the FXAA algorithm is sparse, and much of it has been lost to time. The original FXAA algorithm was published in a whitepaper by NVIDIA and was severely optimized by the author on the suggestions of many mostly anonymous contributors. The latest published version of the algorithm (version 3.11) bears little resemblance to the original and no documentation exists on the changes. The 3.11 version of the algorithm is constructed from a maze of C preprocessor macros, and many different variations of the algorithm are possible based on how the parameter macros are defined.
The implementation of FXAA in the r2 package is a set of GLSL expansions of the public domain Fxaa3_11.h header with a few minor modifications (unused parameter removals). Specifically, the PC algorithm is used, with quality presets (10, 15, 20, 25, 29, 39).