In order to investigate the accuracy and performance implications of using float32, float64, or fixed point arithmetic, a research branch was created.
Supporting float32 and float64 in parallel and comparing them extensively proved somewhat simple with the help of templates. However, fixed point arithmetic differs considerably and instead a playground was created to investigate.
The conclusions can be found in detail here and the code lives here.