I would have wished to see the error analysis section expanded a bit, or maybe seeing some sort of tests to validate the max error. In particular if the mathema

I would have wished to see the error analysis section expanded a bit, or maybe seeing some sort of tests to validate the max error. In particular if the mathematical approximation function arctan* has max error of 1/10000 degrees then I'd naively expect that the float-based implementation to have worse error. Furthermore it's not obvious if additional error could be introduced by the division float atan_input = (swap ? x : y) / (swap ? y : x);