Error Qualification

[Jmeyer1292]

A nice feature of the library is that allows you to use floating point precision (instead of double) if you want the extra bits of speed. The cost is accuracy. This shows up in our tests but I want to make a tool that qualifies it for a set of parameters. It's different for different robots - big ones suffer more from accuracy problems.

[gavanderhoorn]

really just curious here: have you ever tried to quantify the performance difference? I'm aware of the choice between float and double having an impact on things like gpgpu and when working with really large matrices (fi). For the nr of elements in the vector/matrices opw_kinematics deals with, do you observe a similar effect?

[Jmeyer1292]

It's probably overkill to use anything but double. I just wanted the option to use this in things like gpgpu programming.

In my tests (found in open prs) I have seen:
Fk float vs double: 0.2 vs 0.5 us.

Ik float vs double: 1.5 vs 3 us.