Sunday, September 20, 2015

You can talk “the Talk”, but can you loop “the Loop”?

Humans, parrots, donkeys, flies and pigs have that thing called a "heartbeat". X times per seconds it pumps around our blood. Games do the same more or less, with the only difference that an insane high heartbeat will make you drop dead, while games thrive on it. You should try to get the cyclus done at least 30 times per second.

This article is meant for beginners, explaining the GameLoop and more specifically the role of timers in a game. At the end, a small example of the Engine22 eGameLoop component is given.

Get ready for the Launch!

When programming something, there are generally three ways to execute something:
·         Single shot
·         Event driven
·         Looped

Batch files or scripts are usually good examples of a single-shot program. You start the program, a list of instructions is executed one-by-one (eventually pausing to query the operator to do X yes or no), and then it shuts down. Copying or converting files, downloading data, starting an advanced calculation in the background, unpacking files, printing a document, et cetera.

Event Driven programs are your usual Desktop Application. It starts, initializes stuff, and then waits for user-input. Press key, click a button, swipe the screen, drag & drop items, and so on. A Form-based application made in Delphi or .NET are good examples. Components like a button or listview generate all kinds of events when being pressed, entered, changed, moved-over, and so on. On a deeper layer, the OS (Windows, Linux, …) is registering input and sends your application “messages”, which are then translated to events you can chose to use.

By the way, in industrial types of programs, events can also be “interrupts”. Increment counter if sensorX generates a pulse, engage safety procedure when door-sensor loses signal, et cetera. Just saying input doesn’t have to be a mouse-click or keyboard button. A mouse-click or button-press by the way generates an interrupt in the OS as well, causing a message being sent to your application.

C’mon baby Do the Loop

And then we have the “Looped” kind of program. Grandpa telling the same story over and over and over again, and again. Until you knock his head with a wheelchair maybe. You write code, and then tell to repeat this code until X happens. Where X could be the end-of-your-program, a STOP signal, or whenever this (sub)task is considered completed/aborted. Note it executes the same code again and again, but that doesn’t mean the exact same thing has to happen. A typical scenario is that we check input parameters each “cycle” which may alter the routing (if button1Pressed then … else … ) or behaviour of other code parts.

Industrial applications, often running on PLC’s or Micro-Controllers, are usually doing “the loop” approach. Each cycle, they check input(sensors / touchscreen), run regulators, and send output (relays, pumps, valves, servo’s). It’s the heartbeat that keeps the program going. And if there is no heartbeat (stuck in a loop, error raised, hardware issue), the “Watchdog” will time-out and reset the chip.

And games aren’t much different. Note that applications can combine both events and (multiple!) loops, but again the “Game Loop” is what makes your game drawing, moving, sounding and doing stuff continuously.

You probably know how to make an ordinary loop:

            while not ( flagSTOP ) do begin
                                runTask(  ….  );

There is one little-big problem with this code though; no matter how simple “runTask( … )” is, your CPU will go 100% and everything else freezes. It’s trying to execute “runTask” as much as possible, no pausing, no breathing-space for any other application.

This piece of code has two problems. We have no speed-control (the heartbeat pacemaker going mad). And since it’s executed in the main-thread, this means it will block other main-thread tasks in the same program, such as refreshing the screen or handling button-clicks. As for the speed-control, one seemingly simple solution would be using a Timer component. Delphi, .NET, Qt, Java, they all have Timer components that can run code each X milliseconds. The “idle” time between two cycles is then available for other stuff.

Real-time TV

Problem solved then? Hmmm, not quite. At least, it depends on the accuracy we want. See, your default Timer isn’t very accurate. Why? Because the Windows OS (don’t know for Linux) itself doesn’t have a very accurate timing. Why? Because Windows OS isn’t designed as a “Real-time” operating system? Why? Because Windows doesn’t need to be. Why? Because Windows typically isn’t used on time-critical hardware, such as a vehicle ECU or a packaging machine controller. Why? Ah, drop dead.

Though we all love a good performance, 99,9% of the Windows applications isn’t time-critical. People won’t die if we miss a frame, and we don’t control a packaging machine that screws up if its pneumatic valve is powered 10 milliseconds late. Machinery often requires a real-time system to guarantee the same predictable results over and over again (24/7!). Which is why you shouldn’t use a Windows computer for that. Why? I told you why, because the Windows OS isn’t designed for real-time applications. It does not guarantee that your program will execute the loop again over 872 microseconds. On a typical Windows computer, we have dozens of other programs and background tools running at the same time, which can all claim the CPU or memory, so we simply can’t give any guarantees.

Now a game isn’t time-critical in the sense that you will get hurt if the framerate drops from 60 to 30 in all of a sudden. Although… I’ve seen kids on Youtube that went ape after getting shot in a Death-match game due a computer hick-up… Nevertheless, we want a smooth experience. If your television refresh rate fluctuates between 15 to 100 FPS all the time, you’ll be puking after five minutes. Like in machinery, we want to run our game at a certain pace, and guard that interval.

Say we use a Delpti TTimer and put the interval on 16,66 milliseconds. That means the timer-event is triggered 1000/16,66 = 60 times per second (thus FPS = 60). In theory. We’ll use this event to execute our game-loop. In this loop we could:


Now since we have only 16,66 milliseconds between 2 cycles, that is quite a lot to do in a short time!! Yep, true, but you’ll be amazed what a computer can do. Don’t forget about half of the work is done by another processor-set, the video-card, which is on steroids. The picture above implies everything happens in sequence, but it's common to have some multi-threading going on to execute these sub-tasks in parallel. While the video-card is drawing, you could update audio for example. You’ll also learn that a lot of programs out there are slow as shit because bad programming. If your photo-realistic game can poop 40 frames per second, there is no reason why some silly editing program takes 5 seconds to do a single simple task.

A good engine is a powerful beast, doing thousands, no MILLIONS!, of calculations every second. The key to get this fast is not really doing formula’s in a very optimized way, but avoiding stuff that doesn’t have to be calculated (every cycle). Don’t do extensive searches if you could also do that just once during the loading-phase. Use smart algorithms to to reduce searching lengths. Don’t calculate advanced physics for an object miles away. Don’t animate a 30 legged centipede that is somewhere behind you. Anyway, optimizing is another story.

60 Frames per Second is a nice strive, though I’m already happy if the game just keeps above the 30 FPS line. Above thirty, the human eye thinks animations are smooth, but lower framerates will appear choppy. How many frames you can reach, depends on what you’re doing (how complex & how many entities), and how fast the computer is obviously.

If we’re trying to do more than the computer can handle, it means the framerate will drop and our timer will be executed repeating without any delays between the cycles. This can be temporarily, for example when entering a heavy scene, or if the computer is doing a Virus-Scan on the background (slowdowns aren’t always our fault!). Doctors making rush-hour on the ER, more bloody patients coming in than we can handle. If it’s structural, meaning we never reach our desired framerate and the CPU hitting 100% all the time, we should consider lowering the target framerate (allow less patients), doing less code/optimize (train our doctors), or switch to a more powerful platform (a bigger hospital).

Take a break

At the other hand it may happen we can easily perform our tasks in the given timeframe. If the timer runs at 58,8 FPS, our timeframe is 1000/58,8 = 17 milliseconds. If doing our GameLoop only takes 10 milliseconds, we have 7 more “free time” milliseconds. Which is great, because this allows us to do other stuff, it gives some room to other background applications, and otherwise at least the CPU isn’t running 100%, making a lot of noise all the time.

Here is the tricky part though. After you finished the Game-Loop, you should check how much time that took, and how much take you can rest before taking the next step. A Delphi TTimer does that, but not very accurately, because non-real-time Windows isn’t generating timer-pulses very precisely on the background. That also causes our beloved “sleep( )” function to be unreliable. Calling “sleep(1)” may actually put your thread in bed for 10 milliseconds or so. So, how to keep a steady framerate then? 

Engine22 “eMainloop” class

There are several High-Precision timer components for Delphi, and I’m sure the same thing is available for .NET or any other language. Engine22 also provides the “eMainloop” class (E22_Util.Mainloop), which is sort of a timer. You set it up by giving a desired target-framerate -which is basically a maximum speed-, and a Callback function. This callback function is called every time the eMainLoop object triggers. So, typically you execute your game (or whatever it is you’re making) stuff there.

                Looper : eMainloop;

                Procedure TForm1.initialize();
self.looper             := eMainloop.create( self.handle );
self.looper. setTargetSpeed( 60 );
self.looper.setCallback( self. gameMainLoop );
self.looper.enabled := true;

                procedure TForm1.gameMainLoop( const deltaTime : eFloat );
                end; // gameMainLoop

So “gameMainLoop” gets called, 60 times per second (hopefully) in this example. The elapsed time between 2 frames, in theory 1000/60 = 16,66 ms is given as an argument you can use. How to? Check the “DeltaTime” part at the bottom of this article:

How it solves the timing issue? By not using the Windows OS timer messages, but using the Windows vWMTick_ASAP signal instead. This one is given mega-fucking-fast. The application is allowed to process messages every time we receive a tick, and we measure the elapsed time using the Windows QueryPerformanceCounter() function. This function returns an ever incrementing tick-counter, which can be converted to milliseconds by dividing it with the clock frequency, which you can get via QueryPerformanceFrequency( ptrFrequency ). If the elapsed time exceeds our targetframerate, we call the given callback, “gameMainLoop” in this case.

Enough for today. For people with some experience, this whole story probably sounds all too obvious. But for a newcomer, it's probably good to understand what's going on. After all, game-code doesn't quite look like a common (Event driven) Desktop program. And since the looping/timer is such an essential thing, you'd better not rely on the standard Timer and dig a bit deeper in order to get control.

No comments:

Post a Comment