Yes, it drifted a bit after I quantized the coefficients.mystran wrote: The error doesn't look like proper minmax. I would guess this is a Chebychev approximation?
I have checked the instruction timings on https://www.agner.org/optimize/instruction_tables.pdf and I can't verify this claim. Division and sqrt have the same performance across multiple microarchitectures.Division in general is slower than sqrt() actually.