CORDIC

Published on 2016-12-21 in Tote.

Since the flu season began, I’ve been lying in bed sick quite some time. While I don’t recommend doing coding with a fever (takes way too much time to track down the trivial mistakes you make), you do have some free time to explore some coding ideas that you didn’t bother with before. One of such things on my to do list is rewriting Tote’s inverse kinematics code to use only integer math, and I found a marvelous family of algorithms that can let me do it.

If you look at my explanation of Tote’s inverse kinematics at http://tote.readthedocs.io/en/latest/ik.html , you will see that I’m using two trigonometric functions (actually inverse trigonometric functions) in there: atan2 and acos. They both return an angle – in the case of the standard library functions, in form of a number between -PI and PI (or was it 0 and 2PI? I can’t remember). That’s a pretty arbitrary range, that I have to then convert into servo pulse width in the range between 600 and 2400µs. Wouldn’t it me great if they instead returned integers from 0 to 1800, so that I could use them for the servo pulse width directly?

But how are such trigonometric functions calculated? There are several ways. You can approximate them with a polynomial, using a so-called Taylor series. You can also have a lookup table, with some interpolation for the values in between. But both of those approaches involve quite a lot of multiplications and divisions, and possibly floating point (or fixed point) numbers. Which is slow on a microcontroller such as the one used in Tote, not to mention a lot of code being generated (especially for the lookup tables). Fortunately there are the CORDIC algorithms, which involve only a lookup in a small table, some additions and some bit shifts. And of course you do it all on integers. Plus, you have very good control over the accuracy – the more iterations of the algorithm you do, the better results you get.

How does it work? The heart of the algorithm is a small mathematical trick that lets you rotate a point (x, y) around the center of the coordinate system (0, 0) by a fixed angle using just additions and bit shifts, provided you know the arcus tangent value of that fixed angle. So you make a lookup table with a dozen values for 45°, 22.5°, 11.25°, 5.625°, and so on (each angle half of the previous one), and you do a binary search for your value.

Let’s say you want to calculate cos(a). Imagine a length 1 line starting at (0, 0) and going to (x, y), with the angle between the x axis and the line being a. Then the value we are looking for is x. How do we find x? We will start with a line at angle 0° and coordinates (1, 0), and rotate it (using our additions and shifts) by 45° one way or the other, depending if our a is larger or smaller than 0. We will get some point (x1, y1) this way. Then we will rotate it again, by 22.5°, one way or the other depending on where the angle a is compared to our current angle. And we will repeat that, rotating by a smaller angle each time, until we get close enough to out angle a. At that point we have calculated our cos(a) (and also sin(a) at the same time).

Calculating the inverse trigonometric functions is similar, except we start with the line at (x, y), and try to rotate it to be as close to (1, 0) as possible, keeping track of how much we have rotated it so far. Once we got close enough to (1, 0), the total angle is our answer. Simple.

All that is left for me is to actually write some code, test it, tune it to use the smallest types possible and to work with enough accuracy for my particular use case, and put that code in Tote. I will let you know when that happens, but first I need to get over this flu.