Kepler Project

Continuous Integration and Delivery Made Easy

These are not really advanced topics, but I wanted to write about them anyway. So here it is.

Continuous Integration

Sometimes, I report bugs in software. Others, in libraries. Or I can just fix things by myself. And more. Point being: there are lots of bugs everywhere. What do we, software developers, do to avoid writing too many bugs?

We run tests. Unit tests. Integration tests. System tests. Acceptance tests. And where possible, we write software to do these tests for us. Usually, that means unit tests and integration tests. Usually, such tests won’t guarantee that our code is without defect. But at least they give us confidence that it does work for the cases we do tests. If we change something and the tests still pass, we probably have not broken something major. This let us iterate quickly without having to double and triple check whether a change broke one of many features. But even then, starting the tests and waiting for the results is burdensome, so we might get complacent and forget to run them before releasing our changes.

So we get even more software to remember that for us. That’s what we refer to as “continuous integration”, or CI for short. Whenever we submit a new feature, a bugfix, or even if we just want to slightly change the shadow under a button, the CI will run the tests and make sure we have not broken anything. For instance, in the CI of Kepler Project, I run linters to find obvious defects in code, static analyzers to find less obvious defects, as well as a suite of unit tests. I plan to add integration tests at some point, but it is slightly annoying to run them in graphical software.

Continuous Delivery

Once we are confident our changes will not accidentally launch the nukes, we want to actually release them to our users. But software developers are lazy, and they will spend hours automating a task that only takes minutes. In the end, this is the next step after continuous integration, so we call it “continuous delivery” (CD).

An added bonus of a well-designed CD is that you will be able to tell exactly what code was used to create a specific release. For instance, in Kepler Project, the CD starts whenever I create a new git tag. The release it generates uses the tag as the version number. Since a git tag points to a specific state of the code, I can easily check out the code corresponding to a version, to locate a bug, for instance. I can also browse all the changes between two versions.


If you have any kind of non-trivial project, you will want to at least run a CI. And even a basic CD will probably help you in tracking changes from a released piece of software back to the corresponding code in your code history.

Kepler Project

Building for Windows without Running Windows

By the way, I use Linux

1998 2008 2018 2022 will be the year of the Linux desktop. Definitely sometimes before 2028. Maybe before 2038? Ok, maybe not. In any case, I am the most comfortable working (and studying, and procrastinating, and gaming, and procrastinating) on Linux (yeah, yeah, GNU/Linux or GNU+Linux, I know). Actually I am a weakling running Debian with the non-tiling WM Xfce4 (I could as well install Xubuntu at this point).

Anyway, the point is that I still need to target Windows as a platform for the 96 % of users who have not yet seen the light. But I do not really want to have to reboot to switch to a different operating system and maintain multiple developing environment. Besides, I need to build my project in the CI/CD pipeline for automatic testing and release building, but I really do not want to deal with a Windows virtual machine in the cloud.

If only I could build for Windows without running Windows…


Well, it’s a thing. Actually, compiling for X without running on X is a thing. It’s even pretty much mandatory when you cannot run a compiler on the target system. So, if you are developing for Arduino, you are most likely compiling from amd64 (Intel and AMD microprocessors) towards avr (Atmel’s microcontrollers). It’s called cross-compiling.

What makes cross-compiling not completely trivial is that the file hierarchy and dependency management are mostly built around the assumption that compiled code will be run on the current system. So, on my computer, binary libraries in /usr/lib are amd64, /usr/bin/gcc compiles to amd64 and apt installs amd64 binaries by default. Debian does support various architectures, so, I can tell apt to install arm64 libraries, and it will install them in /usr/lib/aarch64-linux-gnu/. Binaries of different architectures, however, will conflict: both amd64 and arm64 Firefox will want to be /usr/bin/firefox. But, in any case, I do not want a gcc compiled for arm64! I want a gcc compiled for amd64 that will compile to arm64. For this, I have to install gcc-aarch64-linux-gnu. With this, I can easily set up a working environment where I will be compiling for arm64 from my amd64 computer.

Cross-Compiling to Windows

But wait, I do not want to cross-compile to arm64! I actually want to cross-compile from amd64 to amd64, but for Windows! Does that even make sense? It does! If I install mingw-w64, I get a new gcc at /usr/bin/x86_64-w64-mingw32-gcc (actually a symbolic link to /usr/bin/x86_64-w64-mingw32-gcc-win32). And then:

$ cat hw.c
#include <stdio.h>
int main(void) {
    puts("Hello World!");
$ x86_64-w64-mingw32-gcc hw.c
$ file a.exe 
a.exe: PE32+ executable (console) x86-64, for MS Windows

I actually do get a PE file, an executable for Windows!

Good, I can compile code for Windows. Now I need to install my project’s dependencies, such as GLFW3. But there is no such thing as libglfw3-mingw-w64-dev. In fact, there are only a few libraries available for this platform in the Debian repositories:

  • libassuan-mingw-w64-dev
  • libgcrypt-mingw-w64-dev
  • libgpg-error-mingw-w64-dev
  • libksba-mingw-w64-dev
  • libz-mingw-w64-dev
  • libnpth-mingw-w64-dev

For everything else, you’ll need to compile the dependencies yourself and install them in /usr/local/x86_64-w64-mingw32/. Yup, no magic wand here. And of course, you will need to cross-compile them to Windows. And fix the bugs you will encounter while doing so.

Pro-tip: write yourself a script to build and install these dependencies. It will serve as documentation for when someone (probably future you) wants to install the development environment. And you will have to do it one way or another if you want to cross-compile in Docker for your CI/CD pipeline.

Running Windows binaries from Linux

How do we run the “Hello World!” example from earlier? Easy!

$ wine a.exe
[useless warnings]
Hello World!

Yup, that’s it.

Okay, okay, it is still useful to test the Windows build on an actual Windows or, at the very least a Windows VM since it sometimes behave differently.

Anyway, that’s all for today. And don’t think too much about floating point determinism when cross-compiling.

Kepler Project

Astronomical Depth Buffer

Guess what. We’re going to talk about floating point arithmetic and what happens when using distances that range from the millimeter to Pluto’s distance to the Sun. Yes, yet again.

Floating Point Arithmetic

OpenGL uses single precision floating point arithmetic. In any case, that’s what your GPU prefers to use. This means you can be as precise as 60 parts in a billion, but not more. That’s more than enough for drawing human-sized objects: the error is much smaller than the width of human hair. And it causes no problem when drawing the Solar-system: the error to Pluto’s orbit is 352 km, which is not visible when looking at the whole thing. But you are going to have issues with drawing a human if you do not know if it’s right here, or 352 km away.

We can avoid the main problem by doing most computations in double precision and centering around the camera before switching to simple precision for the rendering, as I briefly touched in a previous article. And that’s fine! We do not need more than single point precision to locate a pixel on the screen after all.

So far so good. But using a wide range of coordinates in the scene surfaces another issue: the naive use of the depth-buffer.

The Depth-Buffer

3D rendering, at its core, is not very complex. You ask to draw some triangles, your GPU does some math to know where it should draw it on your screen and give colors to the pixels that end up in it. And that’s it! Well, almost. There is one tiny remaining problem.

Let’s say there is a green square close by, and a red triangle farther away (behind the green square). On the left side, this is what we expect to see. On the right side, this is what happens if we draw the green square first: when we draw the red triangle, its pixels will overwrite these of the green square. (source)

Each pixel might be in the direction of various objects. But, most of the time, it should be of the color of the closest object. If we just draw all the objects overwriting the pixels each time, we might show something which is supposed to be hidden!

The most intuitive way to fix this is to just draw the objects in the appropriate order! This is z-sorting. But it is not a trivial task, and will not work in all cases (when A is partially behind B, B partially behind C and C partially behind A).

The more robust way is to remember for each pixel, how far it comes from. When drawing something over this pixel, we check whether the new value is closer. If it is, we use the new value, otherwise we keep the old one. This is the depth-buffer. For each pixel, we’ll track its current color (red, green, blue and alpha), and its current depth.

Precision Issues

By default, OpenGL actually sets the depth-buffer (d) as the inverse of the distance from camera (z): d = 1 / z. This is a secondary effect of the perspective transform, which makes things look actually 3D.

If we divide the range of d in regular interval (on the left), we notice that they cover vastly different intervals in actual depth (on the bottom); this causes precision issues for objects which are far away (source)

There are various ways to fix this. The Outerra blog offers two approach for a logarithmic depth buffer:

  1. explicitly set the depth-buffer from z (or w) in the fragment shader
  2. override z when it will not affect perspective (which relies on w) and let OpenGL use it to set the depth-buffer

The first method works very well, but disables various optimizations, so it slows down rendering a bit. The second method lets OpenGL do its thing — so it is fast — but it does not work well with large triangles close to the camera.

For more information, you can also have a look at another article from the Outerra blog.

Writing this article has forced me to better understand the technique, and to fix a bug due to using the second method incorrectly (I had kept the multiplication by w from an earlier method).

With this out of the way, you can draw things very close-by or extremely far away in a single scene! No need for hacks such as splitting the scene in a close-up part and a far-away part!

Kepler Project

Solving Kepler’s Equation 5 Million Times a Second

I know it is a controversial opinion, but you might need to know where things are when running a simulation. As a bonus, it helps you in knowing what to draw on the screen.

Of course, you can just use a position vector (x, y, z) where x, y and z are the coordinates of the object. But things have the inconvenient tendency not to stay in the same place for all eternity. So we’ll want to know where that object is at the given time.

Now, the more generic approach is to run a physics simulation. We compute the acceleration of the object, update its speed depending on the acceleration and update its position depending on its speed. By the way, use Runge-Kutta 4 for this, not Euler’s method.

Gravity accelerates the Moon towards the Earth, so its speed will tilt counter-clockwise in the picture (source)

That would work for things in orbit too. But we can do much better in this case. This one guy, Johannes Kepler found simpler formulas to describe the movement of two orbiting objects (e.g. the Sun and the Earth, or the Earth and the Moon). Basically, the trajectories will be ellipses, and we can deduce the position of an object in its orbit at any given time.

The orbits of planet1 and planet2 are ellipses such that one focus is the Sun (source)
The area swept by an orbiting object is always the same during a fixed duration (source)

We can describe the position of the object in its orbit as the angle between the direction of the periapsis (where the planet is closest to its sun) and the direction to the object, taken at the relevant focus. This is the true anomaly. Everything is an illusion, so we also like to pretend the ellipse is actually a circle and define the eccentric anomaly. But these angles are not linear with time. So we also define the mean anomaly, which works like an angle (goes from 0 to 2π) but is linear with time. Once we know the direction of the object, we will only need to know its distance, r.

And here are the formulas to get from time to the true anomaly, and how we then obtain the distance from focus (e is the eccentricity of the ellipse):

  • mean anomaly (M): M = M₀ + n × (t – t₀)
  • eccentric anomaly (E): M = E – e × sin E
  • true anomaly (f): $latex \displaystyle f=2\arctan\sqrt{\frac{1+e}{1-e}} \tan\frac E2$
  • distance from focus (r): $latex \displaystyle r=\frac p{1 + e \cos f}$

You won’t be calculating this in your head, but the expressions are pretty easy to translate to working code. Well, except for one. Notice how the eccentric anomaly is not expressed as a function of the mean anomaly? Well, that’s the hard part. And it’s called Kepler’s equation.

Kepler’s equation in all its glory

We do not know a simple way to express E as a formula depending only on e and M. Instead, we must proceed by successive approximations until we decide we are close enough. For this, we will need two things:

  • a starting value E₀ for E
  • a way to improve the current approximation of E

Most introductory books and online articles I have found will tell you to take E₀ = M. This has been obsolete since 1978. It works well enough, but fails in corner cases (near-parabolic orbits around periapsis). And modern starting values are close enough to limit the number of improvements you then need to make. I reviewed 21 papers, selected the 3 most relevant ones, and implemented 21 methods for picking starting values (3 papers for the hyperbolic case with 5 methods for the hyperbolic case).

To improve the starting value, old references will make you do E := M + e sin E, but this is pretty naive. Most modern references will use Newton’s method, which helps you solve an equation faster as long as you know its first derivative. But you can go further by using the second derivative using Halley’s method. You can go arbitrarily far using Householder’s method.

Great, now I have 21 × 4 = 84 approaches to try (5 × 3 = 15 for the hyperbolic case). Which one is the most accurate? Which one is the fastest? And, more importantly, can I find a fast and accurate approach?

You can spend years studying the theoretical properties of these functions. Or, we just try them for real. It is the most efficient way to know how they actually behave on real hardware with real implementations of numbers.

I wrote a benchmark tool to test both the accuracy and the speed of each approach. It turns out that only a handful of approaches are accurate enough for my taste:

# Elliptical case
## Naive method
## Newton's method
## Halley's method
     580 c  3.580e-16  gooding_10
     764 c  3.926e-16  gooding_11
     577 c  3.604e-16  mikkola_1
     742 c  3.653e-16  mikkola_2
## Householder's third order method
     606 c  3.645e-16  gooding_10
     809 c  3.623e-16  gooding_11
     641 c  3.935e-16  mikkola_1
     764 c  3.904e-16  mikkola_2

# Hyperbolic case
## Newton's method
## Halley's method
    1454 c  3.731e-15  mikkola_1
    1578 c  3.725e-15  mikkola_2
## Householder's third order method
    1704 c  3.618e-15  mikkola_1
    1757 c  3.618e-15  mikkola_2

The first column is the number of CPU cycles per solve of Kepler’s equation. The second column is the worst relative error over the test set. The third column is a short name for the method for picking a starting value for E.

As we can see above, no approach seems to work well enough with the naive improvement, or with Newton’s method. And using third derivatives does not do much. gooding_11 and mikkola_2 are relatively slow. This leaves gooding_10 and mikkola_1. I picked mikkola_1 because the paper also treated the hyperbolic case. It comes from A cubic approximation for Kepler’s equation by Seppo Mikkola.

And after all this, I have a very robust way to know where stuff is at any time! Of course, you should not forget to take care of floating point precision issues when implementing all of this…

Appendix: References

Appendix: Full benchmark

# Elliptical case
## Naive method
     264 c  2.500e-01  smith_1
     276 c  5.303e-01  smith_2
     360 c  2.000e-01  smith_3
     444 c  9.701e-02  smith_4
     385 c  1.667e-01  smith_5
     480 c  6.495e+01  smith_6
     246 c  2.500e-01  gooding_1
     299 c  2.000e-01  gooding_2
     356 c  1.667e-01  gooding_3
     246 c  5.303e-01  gooding_4
     460 c  9.701e-02  gooding_5
     253 c  2.417e-01  gooding_6
     325 c  7.497e-02  gooding_7
     334 c  7.497e-02  gooding_7b
     378 c  1.277e-01  gooding_8
     423 c  7.499e-02  gooding_9
     455 c  2.376e-01  gooding_10
     608 c  8.020e-02  gooding_11
     286 c  1.439e-01  gooding_12
     448 c  4.342e-01  mikkola_1
     629 c  4.426e-01  mikkola_2
## Newton's method
     522 c  8.517e-03  smith_1
     539 c  1.379e-02  smith_2
     578 c  8.111e-03  smith_3
     697 c  8.111e-03  smith_4
     706 c  8.111e-03  smith_5
     900 c  1.028e+02  smith_6
     481 c  8.517e-03  gooding_1
     583 c  8.111e-03  gooding_2
     561 c  8.111e-03  gooding_3
     476 c  1.379e-02  gooding_4
     686 c  8.111e-03  gooding_5
     537 c  3.528e-02  gooding_6
     536 c  1.378e-02  gooding_7
     585 c  1.378e-02  gooding_7b
     702 c  2.524e-04  gooding_8
     680 c  1.379e-02  gooding_9
     731 c  1.117e-10  gooding_10
     940 c  2.072e-08  gooding_11
     528 c  5.673e-03  gooding_12
     697 c  1.825e-06  mikkola_1
     815 c  1.825e-06  mikkola_2
## Halley's method
     395 c  1.250e-01  smith_1
     419 c  2.553e-03  smith_2
     488 c  6.250e-02  smith_3
     596 c  1.710e-02  smith_4
     557 c  4.167e-02  smith_5
     989 c  5.130e+01  smith_6
     429 c  1.250e-01  gooding_1
     503 c  6.250e-02  gooding_2
     524 c  4.167e-02  gooding_3
     426 c  2.553e-03  gooding_4
     587 c  1.710e-02  gooding_5
     410 c  6.127e-03  gooding_6
     467 c  2.548e-03  gooding_7
     440 c  2.548e-03  gooding_7b
     508 c  9.153e-05  gooding_8
     552 c  2.550e-03  gooding_9
     598 c  3.580e-16  gooding_10
     821 c  3.926e-16  gooding_11
     430 c  1.028e-03  gooding_12
     594 c  3.604e-16  mikkola_1
     742 c  3.653e-16  mikkola_2
## Householder's third order method
     438 c  1.562e-02  smith_1
     439 c  6.779e-04  smith_2
     501 c  7.812e-03  smith_3
     667 c  2.138e-03  smith_4
     533 c  5.208e-03  smith_5
    1053 c  6.535e+01  smith_6
     453 c  1.562e-02  gooding_1
     501 c  7.812e-03  gooding_2
     594 c  5.208e-03  gooding_3
     429 c  6.779e-04  gooding_4
     595 c  2.138e-03  gooding_5
     454 c  1.638e-03  gooding_6
     526 c  6.736e-04  gooding_7
     585 c  6.736e-04  gooding_7b
     583 c  9.493e-09  gooding_8
     652 c  6.741e-04  gooding_9
     681 c  3.645e-16  gooding_10
     846 c  3.623e-16  gooding_11
     486 c  2.724e-04  gooding_12
     651 c  3.935e-16  mikkola_1
     752 c  3.904e-16  mikkola_2

# Hyperbolic case
## Newton's method
     780 c  7.245e+25  basic
    1064 c  1.421e-03  mikkola_1
    1201 c  1.970e-04  mikkola_2
    1014 c  8.559e-01  gooding
     892 c  1.515e+19  pfelger
## Halley's method
    1134 c  1.729e+17  basic
    1396 c  3.731e-15  mikkola_1
    1526 c  3.725e-15  mikkola_2
    1265 c  9.160e+10  gooding
    1237 c  1.791e+13  pfelger
## Householder's third order method
    1260 c  1.323e+07  basic
    1631 c  3.618e-15  mikkola_1
    1595 c  3.618e-15  mikkola_2
    1467 c  2.276e+17  gooding
    1365 c  9.997e+01  pfelger
Kepler Project

Floating Point Woes in Three Dimensions

Sometimes, stuff in space rotates[citation needed]. In this article, I am going to explain the basic idea to simulate rotation efficiently, and a few technical hurdles I encountered while implementing it.

I have not yet got down to reinventing the wheel writing a physics engine. Besides, I don’t want to simulate physics just to keep track of the orientation of the object through time. Doing so would drastically limit how much we can time-wrap (to maybe ×10,000).

Instead, we can just assume a constant rotation (angular velocity), and deduce the orientation at any given time from the base orientation and the elapsed time. In other words, we set:

Orientation(t) = Orientation(t₀) + AngularVelocity × (t – t₀).

This approach fits very well with planets, whose angular velocity can be safely assumed to be constant.

To handle simple changes in angular velocity, we can just update t₀. So, when the ship thrusts its RCS to pivot, we set t₀ to the current time, Orientation(t₀) to the current orientation, and update the angular velocity.

Of course, the assumptions we make are not very accurate for non-spherical objects. For instance, we will not observe the Dzhanibekov effect (also known as tennis racket theorem) with this approach.

One thing remaining: how do we represent an orientation or an angular velocity? There are essentially three ways:

  • Euler angles
  • axis-angle
  • rotation matrix
  • quaternion

Let’s go over each in more detail.

Euler angles use three angles around fixed axes. They are somewhat intuitive, but are very inconvenient to work with.


You can also find the rotation’s axis to get the axis-angle representation. You’ll need three values to describe the axis, and one for the angle.


Rotation matrices can be easily combined. They also adapt well to represent translation and scaling, when using homogeneous coordinates (convenient for 3D rendering). But they need at least 9 parameters.


The theory behind quaternions is hard to grasp, but you do not need to. What matters is that you can easily combine them, extrapolate them, or interpolate between two orientations, and they only use four values.

Actually, an intuitive way to think about them is that the first components x, y and z correspond to the axis of rotation, while the w component represents the angle of the rotation: if your angle of rotation is θ, then w=cos(θ/2). We just need to keep their norm to 1, so we scale x, y and z appropriately.


So, clearly, I use quaternions wherever I can. But there is one thing to keep in mind. Quaternions (almost) uniquely represent a rotation, and a rotation by angle x is the same as the rotation by angle 2π + x around the same axis. So, if we try to store an angular velocity of 2π rad/s (one rotation per second) as a quaternion, we’ll just have a quaternion representing an angular velocity of 0 rad/s (well, 0 rotation per second).

The usual way around this is to represent angular velocity (and impulse) as axis-angle, since we can store whatever we want in the angle. But I like edge cases and floating point Eldritch horrors. So, instead, I just scale the angle down by something like K=2¹⁰⁰. We will be fine as long as the angular velocity is smaller than 7.96×10³⁰ rad/s (26 billion billion (no typo) times faster than the fastest spinning object ever).

To summarize so far, we’ll use a quaternion to represent both the rotation and the angular velocity, except that we will scale the angle of the angular velocity by some gigantic value K.

It means we will create quaternions with very small angles.

It means that the component w will be extremely close to 1 (the distance will generally be about 1 / K = 2⁻¹⁰⁰).

It means that the component w will just be equal to 1 (since even double precision floating point numbers won’t have enough precision to tell the difference). We just lost all our precision!

It means that, when retrieving the angle from w (needed when scaling by t – t₀), we will just get 0.

It means that we will multiply 0 by K.

And 0 multiplied by a finite value (even very large) is still zero.

So nothing will ever spin. Great… back to Euler angles I guess?

Not so fast! Although we cannot retrieve small angles from the w component, the information is actually present in the other ones, x, y and z!

But you just told us that the x, y and z components are just the axis of rotation! Did you lie to us? Must my dreams always come crashing done at the altar of floating point numbers. I knew I should have gone with Cthulhu.

A sensible person when confronted with floating point arithmetic

It is true! But an axis is just a direction. When these three value also store a magnitude. Remember when I told you that we had to scale x, y and z so that the norm of the quaternion stays equal to 1? Well, that’s convenient, because it means that √(x² + y² + z²) = sin(θ/2).

And sin(θ/2) is very precise around zero in floating point arithmetic!

Which means we don’t lose any precision!

After checking with Herbie and opening a merge request in GLM (the library I use for manipulating quaternions), I am able to represent angular velocity as quaternions!

“And yet it moves!”

Could I just have used the axis-angle representation for angular velocity? Probably? Would it have been fun? Probably not as much.

Kepler Project

Looking Up-Close at Orbits

In the previous post, I explained the basic idea that is used to draw orbits using line segments, and a little trick to make even very elliptical orbit look good. So, this is where we are:

Looks good.

But if we zoom in:

The line segments are visible again!

What is going on?! I thought we had taken care of spreading the line segments to hide the angle as much as possible!

A distraught hypothetical reader

Well, true. But we did it assuming we were looking straight down at the ellipse. Here, perspective comes into play. Although it conserves lines, perspective has the nasty habits of effecting angles. Here, it magnifies the angles between segments around the point where the orbits appear to bend towards the left. On the contrary, the segments on the left part look pretty much aligned.

But it’s okay. We only need to add a few more points close to the camera to compensate for this. It does mean that they will move as the object follows its orbit.

Here, we can clearly see the additional points moving alongside the focused object on the orbit. When we zoom in, they make sure we don’t see any unsightly corner.

But wait, there is more!

The Earth not being on its own orbit might be a bit of a problem

This is not the path of the Earth-Moon barycenter, which is the surface of the Earth anyway. The line really should pass through the center of the Earth.

So, what is going on? We do have a nice smooth-looking orbit, but it is still an approximation. It just so happens that we have looked at a time when the Earth was in the middle of a line segment. In the diagram below, you can imagine that the line segment is BX and the Earth is betwen the two.


This becomes more apparent as we speed things up:

The Earth follows the actual path of the ellipse while the orbit is just a bunch of line segments. The points between line segments are in the right position, which corresponds to the times when the orbit actually appear to go through Earth’s center.

To avoid this, we just need to ensure that at least one point of our orbit is always in the middle of the Earth. Actually, we can just make it so that it is one of the additional points from earlier!

Now everything looks smooth and the Earth is always on its orbit!

Actually, that was the easy part. Now let’s zoom a bit more for the fun one:

Showing Pluto here, as the effect is more visible. But we cannot ignore it just because it’s not a planet anymore! And the effect is even more visible on smaller objects such as spaceships, because we can zoom closer to the orbit.

To understand what is going on, we need to dive into how we actually calculate the positions of the points of our orbit. Digital computers only know about discrete values (that’s what the “digital” mean). The basic unit is the “bit” (0 or 1) and, by using several bits, we can count natural integers fairly easily (0, 1, 10, 11, 100, 101, 110, 111, 1000). Add a sign bit, and we’ve got relative integers. We’re not going to be able to represent arbitrary real numbers short of using symbolic computation, but we can at least get something good enough by using decimals. The idea is that we would use two relative numbers: one for the number’s digit, and one to tell where the decimal point should be. And that’s exactly what computers do. Well almost.

Every explanation about floating point arithmetic must start with this example. It is known.

Technically, computers use binary instead of base ten, but the principle is the same. If we really want to use base 10, we can, albeit it’s slower:

But the important point to keep in mind when working with non-integer values with computers is that they are (usually) only approximations of the number we want.

For instance, the approximation of π used by Python is slightly smaller than the actual value of π. So, if we compute the sine of this approximation, we get something slightly bigger than zero.

We can get a good idea of the approximation. There are two usual formats used to represent non-integer: binary32 and binary64, more commonly known as float and double respectively (although Python’s float is binary64). The relative error of a binary32 is roughly 2⁻²⁴ and the relative error of a binary64 is roughly 2⁻⁵³.

The error in binary32 when dealing with values around the radius of Pluto’s orbit is of 352 km (Pluto’s radius is 1,188 km). With binary64, it’s less than a millimeter.

Rendering to a screen which is no more than 8,000 pixels a size should not require more than binary32. So this is what is normally used for rendering. Using binary64 here would degrade performance a lot. But it does mean that we are getting imprecision when drawing things naively.

Let’s say the Sun is the center of our system of coordinates, and we use SI units. Then, the Pluto is 5.9 Tm from the Sun. For rendering, we are going to compute the positions of every object relative to the Sun, compute the position of the camera (our point of view) relative to the Sun, and subtract the second from the first to find the position of the objects relatively to the camera.

Since this is naturally done during the rendering in the GPU shaders, we use binary32. The error from doing computations in binary32 is roughly 352 km in this case. So, let’s start from the camera position; we are at a distance of 0 (exactly). Then, let’s move to the Sun; we are at a distance of 5.9 Tm±352km (relatively precise). And let’s move back to Pluto; we are at distance of (5.9 Tm±352km) – (5.9 Tm±352km) = ±352 km (not very precise). This is called loss of significance.

We really want to avoid doing such a computation in binary32. We must either switch to using binary64 in the rendering, or avoid doing this computation in the shaders entirely. We really want to avoid doing the first one, since it can slow down the rendering significantly, and not all GPUs even support it.

How do we avoid doing this computation in the shaders then? The most intuitive approach seems to switch the system of coordinate to take Pluto as center. But this has issues too, since we still have the exact same imprecision when switch from one coordinate to the other. The solution I opted for is actually simple: I do the imprecise computation using binary64 outside the shaders, which lets me have millimeter precision around Pluto. It can be seen as switching coordinates, but only for part of the rendering.

The loss of significance is still here. But it’s now only 1 mm.

Note: I use the same technique when drawing celestial bodies. This is actually why the object being observed is well positioned at the center of the screen and do not move erratically (unlike in the example below).

This is not all roses however. For instance, if we look at Pluto from the Earth.

This requires a FOV of 36 hundreth of an arcsecond. Current telescope are far from being able to do this, but a software simulation should.

Since the camera is positionned at the center of the Earth, the distance to Pluto is of the order of 6 Tm. With this level of zoom however, we can distinctly see the imprecision of roughly 350 km, causing the dwarf planet to not be perfectly centered in the screen.

Changing the system of coordinates is not quite enough, since the camera is not close to the observed object. So I will need ot find another workaround.

So, seriously, don’t let the floating point bugs bite.

Kepler Project

Drawing Nice Orbits

So, I have been working on Kepler Project again, and I thought I would write about some of the interesting challenges I have encountered and how I overcame them.

The Shapes of an Orbit

Today, we will talk about drawing orbits. I will talk about the orbital computations themselves another time.

Here, we can see the full orbits of the four terrestrial (i.e. rocky) planets: from interior to exterior, Mercury, Venus, Earth and Mars. In the middle, of course, is the Sun, and the outermost orbit partially visible here is that of Ceres, a dwarf planet whose orbit lies between those of Mars and Jupiter. The red circle show the current (as of 2021-01-27) location of the planet in their orbits.

The planets of the Solar System follow an ellipse as they revolve around the Sun. They are fairly close to a circle, except for the notable exception that is Mercury: in the screenshot above, you can tell that the Sun is not in the middle of Mercury’s orbit (the innermost one). In fact, the trajectory of Mercury draws an ellipse whose focus is the Sun.

Zooming out a bit, we see that there are quite a few dwarf planets orbiting the Sun from far away a bit haphazardously. The four regular orbits in the middle are those of the gas giants: from interior to exterior, Jupiter (whose Orbit is partially hidden behind all the red circles of the inner planets), Saturn, Uranus and Neptune. We find the dwarf planet Eris the bottom-right corner, and Sedna, a bit above, whose orbit is so extended, it takes ten thousand years (Earth revolutions) to complete a full revolution in its orbit.

From the first image, and from the usual depictions of the Solar System, you could think that the orbits are always neatly arranged. But that this becomes increasingly inaccurate as we pull farther away from the Sun.

In this side-view, we can better appreciate that the orbits of the eight planets and Ceres lie almost in the same plane, while the orbits of the outer dwarf planets have a much more pronounced inclination.

From this, we know we are going to need to draw ellipses shapes (eccentricity) and sizes (I use the periapsis to characterize the size of the orbits) in all kinds of orientations (since there are three spatial dimensions, we will need three parameters; I use longitude of the ascending node, inclination and argument of periapsis).

Drawing an Ellipse

GPUs let you draw points, line segments and triangles. And that’s pretty much it. The point is that these primitives let you draw anything if you use enough of them. By restricting ourselves to the simplest primitives, we let the GPU designers focus on doing one thing well. Of course, there is a huge swath of features to help you draw millions of them efficiently, but everything you see in a 3D game is pretty much drawn with textured triangles.

So, how do we draw an ellipse when we can only draw line segments? Well, you pick points on the ellipse and connect them together to make a polygon that approximates the real curve:

Here, with only 8 segments to draw an orbit, you might just tell something is off. By the way, notice that, as a consequence, the objects (in the red circles) are not on the orbit we’ve drawn; it will come up later.

If we use enough segments, the illusion becomes good enough:

In Kepler Project, each orbit is drawn using 256 segments. It could become visible at a high enough resolution, but it does the job.

So, we’re done, right? Not quite. We have only looked as near-circular orbits. Look at what happens with a more eccentric (more elliptical) orbit:

I have set the apogee to 924 Mm, at the edge of Earth’s sphere of influence. The orbit has an eccentricity of 0.985.

Drawing a Nice Ellipse

So far so good. But let’s zoom-in closer to Earth and the perigee of our orbit:

The line segments are clearly visible on the right. It would only get worse with more eccentric orbits.

This is not good enough for me. I might accept glitches are extreme and unrealistic orbits, but this is not the case here. We could of course raise the number of segments used to draw the ellipse, and it would work (here, 1024 seems enough). But we should have ourselves why. Why do we need to use more segments in this case? Let’s have a look:

I switched to only drawing the points (well, big points), and only 64 of them per orbit.

The curve is pretty tame in the middle of the ellipse (on the left of the image), but much stronger closer to the extremities (in the center of the image). From this, we should notice that we have more than enough points in the middle, and not enough at the extremities. If we distributed them better, we would not need more points!

How do we do that?

If found a suggestion on Stack Overflow. But the reasoning is not very clear to follow, so I expose it here. We want to make sure the angle between two consecutive line segments is small enough so that it is not visible.

First trick: We can simplify things by reframing the problem as making sure the angle between two consecutive segments is always the same. To understand why it makes sense, imagine all the angle between consecutive segments is always the same; then, if you decrease the angle between two consecutive segments, you have to increase at least one angle between two other consecutive segments (for a fixed number of segments of course).

Second trick: to avoid the annoying part of finding a formula for the angle between two segments depending on the placement of their extremities, we will actually look at the tangents to the ellipse in the points we choose.

Third trick: we already know we want a constant angle difference between two consecutive tangents, but that means we already pretty much know this angle: 2 π / n where n is the number of segments (note: strictly, we should not assume one of the points will have an angle of 0; but we’re going to want to have a point at the periapsis anyway).

These tricks allow us to forget entirely about the line segments and just look at the tangent of the orbit in a given point:

The primary is one of the foci of the ellipse. The center of the ellipse, though, does not map to anything in particular and could very well be in outer space. E is short of “eccentric anomaly”, which is the angle between the periapsis and the object measured at the center (hence “eccentric”). Another common angle is the “true anomaly”, which is the angle between the periapsis and the object, measure at the primary, usually the actual point of reference (hence “true”).

In the figure above, we know α, but we need to know E. Once we know E, the actual coordinates of A can be calculated using the parametric formula for the ellipse: x = a cos E and y = b sin E. And now we use:

Just kidding, we use applied mathematics witchcraft:

$latex \begin{aligned}\tan \alpha &= \frac {\mathop{dy}} {\mathop{dx}} = \frac { \frac {\mathop{dy}} {\mathop{dE}} } { \frac {\mathop{dx}} {\mathop{dE}} } = \frac { \frac {\mathop{d}} {\mathop{dE}} b \sin E } { \frac {\mathop{d}} {\mathop{dE}} a \cos E } \\ &= \frac { b \cos E } { – a \sin E } = – \frac b { a \tan E } \end{aligned}&s=4$

Now, solving for E, we get:

$latex E = – \arctan \frac b { a \tan \alpha }&s=4$

This formula is not convenient because of the discontinuity around α = 0, so we use some more trigonometric sorcery to get to:

$latex E = \frac \pi 2 – \arctan \frac a b \tan \alpha&s=4$

Let’s look at the effect of using this formula on the repartition of our points:

Still with 64 points. They are much more concentrated around the perigee (and apogee on the other side, not visible here)

And the final result when switching back to line segments:

Still with 256 segments

Now, that’s good enough. Actually, it works well even with extreme orbits:

The orbit shown here as a perihelion of 1 Gm and an apohelion of 1 Zm (or about a hundred thousand light-years). Drawing it with 256 line segments looks fine from above, but I actually increase the number to 4096 in this case so that looking at the apohelion when near the perihelion does not look too weird due to perspective (the formula above does not take it into account).

To be continued

The formula above can easily be adapted to the parabolic and hyperbolic orbits.

But there are actually still a few issues when we look up-close (when we actually want to look at our spaceship for instance). But there is enough to talk about for another article. Until then… don’t let the floating point bugs bite!