OpenGL Programming on Linux

This article is not intended to be an OpenGL tutorial or introduction. There are people far more competent in this area than myself, and they have written a number of articles and even books about this subject. Also, even though the project discussed in this article was written and built mostly using Xinside's OpenGL, it is not my intention to discuss the superiority of any of the Linux OpenGL ports or implementations over another. This article tells a cool story about using Linux, and I thought it was worth contributing to the Linux community.

“And, thus,” the professor concluded, “Instead of working on perfecting our home-made ray-tracer—something we have been doing for years—this semester, we'll start something new. You will have to build a 3D, network-capable tank game using OpenGL. You'll be in teams of two or three students. For that, you'll use the RS/6000 workstations we have here, and we'll give you an introduction to OpenGL.”

My ears! They could not believe what I had just heard. As a student in Computer Engineering at the Polytechnical School of Montréal and a computer graphics fan with a good background in ray-tracing, I had been waiting for years to be advanced enough in my studies to be able to take that course in advanced computers graphics. That semester I had finally been able to take the difficult 4th year course, and I had just heard, to my immense disappointment, that instead of working with a ray-tracer and producing high-quality ray-traced pictures, I would have to work on OpenGL. My morale was not high, and fear was making its way along my stomach as I realized I knew a lot less about OpenGL than about ray-tracing. But as one LinuxDoom aficionado would put it: “Armed solely with my Linux Box and my OpenGL Beta product, I plunged into the hostile mass of GL intrinsics, prepared to fight with every last GLfloat variable I had.”

More seriously, OpenGL is a graphics library designed from the start as a hardware-independent interface to be implemented on many different platforms. It uses a client-server approach, similar to the X client-server approach, to provide display of graphics primitives on the chosen windowing system. The server sends commands to the client, and the client displays them.

On X-capable Unix workstations, OpenGL has an extension to the X server named GLX. You can run your OpenGL program on one computer and display it on another, but it requires that the server machine has the needed OpenGL libraries and that the client has the GLX extension. Since those two packages usually come together, this means that both the server and client must be “OpenGL-capable”.

In short, OpenGL is capable of displaying simple geometric objects, showing orthogonal and perspective projections, performing back face removal, doing shading and anti-aliasing, and applying textures to objects. If you want to do something complex—like display a car or a plane—you have to build those objects yourself, and use OpenGL to render them the way you like.

On Linux, to my knowledge, you have the choice of several commercial implementations and one free implementation:

My personal experience was that the product I was using (Xinside's OpenGL second beta, and later, the final product, which was even faster) was of very high quality. It was faster and more compatible than Mesa. Speaking about commercial applications running on a free operating system is a sensitive and slippery issue, especially when there are freely available equivalents (Mesa) and even more so when you happen to find yourself very (or at least more) satisfied by a commercial tool. I found Mesa to be an impressive piece of software, but Xinside's OpenGL beta was noticeably faster and more OpenGL-compatible, since it is a true OpenGL implementation.

So, here I was, a few days later, in front of an RS/6000 workstation, writing the first few lines of code of that soon-to-be tank game and wondering if it was going to run on my Linux box. You see, I had subscribed to Xinside's OpenGL beta program a few months before as a means to remotely run OpenInventor from my Linux box, and thus, I found myself with an OpenGL-capable Linux computer. Later that same day I went home—while my Linux box was retrieving the sample code by FTP—and got ready to compile it under Linux.

The project we were building was using a freely available auxiliary library named libaux. “Fine,” I thought, and I FTPed its source code from the RS/6000 lab and compiled it on my Linux box. It's also available from ftp.sgi.com under the OpenGL sub-directory, along with all the examples from the OpenGL programming guide. With a lot of hope and increasing excitment I got ready to start the sample code and...it crashed, generating a panic file and killing the X server.

The team later figured that this problem was caused by a small bug in the Beta OpenGL release I was using which caused it to misbehave when using a color-indexed color mode and single-buffering. The program, however, ran fine as soon as I switched to use RGBA (for Red, Green, Blue and Alpha) color mode—it even ran slightly faster than on the older RS/6000 workstations we were using!

Granted, those RS/6000 were basic entry-level workstations, and their age (about two years), combined with poor 3D hardware accelerated video cards, proved they were no real match for my P133 with its Matrox Millennium (although OpenGL on Linux only provided software 3D acceleration). For someone who has been used to “This hot stuff runs on workstations—W-o-r-k-S-t-a-t-i-o-n-s—don't even think about running it on your home PC!” this OpenGL on Linux experience was like a dream come true.

The days went by, and we started incorporating more and more code into the project. My team-members had more course work than I did, so I found myself leading the team—writing most of the code in the first part of the project and all of it in the second and third parts. Of course, I was writing it all on Linux—but always verifying later that it ran on the RS/6000 workstations (Murphy, you know?).

Of course, that did not go unnoticed, and some of the students in the class started exploring ways to build and develop their own projects at home on their PCs using NT or that OS-with-an-expiration-date-in-its-name (Windows 95). Others followed my advice that it would probably be better to use a Unix because of portability problems (I thought...er envisioned...er imagined that the Win32 API could be quite different from that of most Unices) and got Mesa running. After all, if you have the choice and if you can do the same things you do at your university at home, would you rather spend nights in a freezing-cold computer lab with armless wooden chairs or work on your home computer?

Problems started to appear just a few weeks after that when we were required to implement and use a timer within the game. That was the first blow for the NT/95 people because, unless you're familiar with the Windows API or have some sample source code, changing Unix's gettimeofday() to a Windows API call is not trivial. After all, if your virtual tank is going at 10 m/s, it should do so no matter what hardware you have, be it a 16 CPU SGI workstation or a poor 80486. Some people got tired of putting #ifdefs and #ifndefs in their code and decided to spend nights in the lab instead.

Then came the network daemon. The idea (mostly at my suggestion) was that the game client running on a particular workstation would fork() a daemon at initialization. The daemon would share one or more memory segments with the client and would have the task of listening on certain ports for broadcast messages sent by other possible network players. Needless to say, these Unix intrinsics marked the end of the Windows port; even if you could run a part of the 3D engine on Win32, you'd still have to do all the network and final debugging on the RS/6000 workstations at school.

But during all this time there was at least one happy Linux user who did not change a single line of code when sending it from his home Linux box to the Risc workstations. And the only time he actually had to put an #ifdef was when the endianness difference between the Pentium and the RS/6000 processor started to show in the byte ordering of the TARGA files he was loading and using for textures. Rumour even has it that he debugged his network code without actually entering the computer lab: in the darkest hours of the night he used two workstations to run his program on and exported the display to his Linux box (which was slow, but functional enough to track down some bugs).

Speaking about performance in OpenGL is, for those of us who don't use a middle to high-end SGI workstation at home or at work, about as important as speaking about OpenGL itself. In our case, around the middle of the semester, it became obvious to both professors and students alike that the RS/6000 workstations we were using were not fast enough for what we were doing with them.

Eventually we switched to another lab of RS/6000 workstations which belonged to the Mechanical Engineering Department. People there ran CATIA—like AutoCad but with ten times the features and the memory requirements. Those workstations were still not inherently faster than a good Linux Pentium PC with enough RAM; most tests, gcc, xv, etc, showed my P133 was about 50-60% faster doing generic operations. But their hardware-accelerated OpenGL graphics allowed my game to run on them at 25 frames per second with 512x384 pixels. By comparison, I was getting a maximum of only 9-10 frames per second with 320x240 pixels on my Linux box, where OpenGL rendering was done by software alone on the main CPU and FPU.

The program still ran, and it was fast enough to allow me to work out most of the bugs and implement new features, but I would personally have enjoyed it a lot more if the Linux OpenGL port I was using had been able to take advantage of the 3D features on my video card to make my programs run even faster. On my end, I tried removing as much un-optimized stuff as possible from the game's main loop to make it run as fast as possible on all platforms.

Here are some stats about the project:

The project, for all the work-teams in our course, is now finished as far as it involved students working on a programming project whose results professors would evaluate. I finished the final “product” pretty much alone and about a week before all the other teams. In the end, our game client probably had the greatest number of features, the most complex graphics, the nicest explosions and the most reliable motion engine—and we got the highest marks possible on the final evaluation by the professors. Somehow, I think that if I had not been able to run everything on my home Linux machine, and do everything when I wanted it and how I wanted it, I probably would not have reached this level of achievement.

Other than showing that some Computer Engineering students are definitely more productive on their home machines than on most computers you can give them access to, this somewhat extraordinary adventure shows that some fields—which, until now, were reserved for high-end workstations—can be explored with something as simple as a good Linux box and some relevant software.

UPDATE: 11/29/2006

If you want to see the pictures and code relating ot this article go to this link: www.step.polymtl.ca/~coyote/graphics_tank.html