Cowboy Programming

Game Development and General Hacking by the Old West

May 30th, 2008

Measuring Responsiveness in Video Games

In this article I suggest that the specifications of a video game should always include a measure called “response time” (also called “lag”, “controller lag”, or “input latency”). Response time is defined as the time between the player using the controller, and the results appearing on the screen. Example: Pressing the trigger button on the controller fires a gun on the screen. Video game response time can be measured with a cheap digital camera, and I explain how.

THE PROBLEM

The “feel” of a game is in large part described in terms of how “responsive” it is. Very often a game will be described as “laggy” or “sluggish”, and by contrast other games will be “tight” or “fast”. I have previously described the technical reasons behind games lacking responsiveness, but I offered no way of measuring the response time, and so the developers have to rely on their own assumptions about the way they read the controller and present the results, and combine that with the subjective assessments of the test department.

Having an accurate way of measuring response time allows the developer to both verify their own assumptions (hence detecting bugs that are adding to the response time), and to provide an objective reference to the claims of the testers regarding the “tightness” of the game. Perceptions of changes in small variables like response time can vary by individual, and being able to measure it objectively will allow you to see if it has actually changed, and by how much.

Game developers also have to make the decision of whether to go with 60fps or 30fps. 60fps will generally have half the response time of 30fps, which can be a deciding factor (along with the smoother motion, which is visually more appealing on fast moving games). However for some games there are other factors that influence the response time. Having an accurate way of measuring the response time allows the developer to more accurately and objectively make a decision on if 60fps is necessary, or if they simply need to tighten up their 30fps game.

MEASURING RESPONSE TIME

Measuring response time is very simple, and consists merely of videoing the screen and the controller at the same time with a video camera that records at 60 fps, and then playing this back and counting the frames between button press and the screen response.

The camera I use is an old model Canon Powershot SD800IS, a relatively cheap camera. You need a camera that supports 60 fps (frames per second) recording. This may be listed as “Sports Mode”, or “Fast Frame Rate”. I believe many of the current line of Canon Powershot SD cameras have this feature, including the popular and cheap Powershot SD770. This is a very popular brand of camera, so it is highly likely that someone on your team will own one.

Set the camera to movie mode, change to to the appropriate mode for 60fps (press the “func” button on the Canon SD and select “Fast Frame Rate”).

Then set it up to point at the television. You don’t need to get the whole screen in shot, just enough so you can see the game actions that will result from your button presses. Make sure that the controller is roughly in focus and hold it angled so that the button presses can clearly be seen.

CALIBRATION AND MEASURING

Plasma and LCD televisions introduce lag of their own, as they frequently do some additional processing to a picture before they display it. This can be as much as a few 60ths of a second, which can be quite significant. Ideally you would do your testing on a CRT Television, which should have no additional lag (the cheaper the better). Your flat panel TV might also have a “game mode” which will reduce (but not eliminate) the lag.

To find how much lag your TV adds, you need to compare an easily measured response with the same response on a CRT TV, or with a known value. To simplify things for you I’ve measured the response time of the PS3 system menus, which come out at a very solid 3/60ths on a CRT, and 5/60th on the plasma TV shown above. Hence I need to subtract 2/60th of a second to get an accurate measure of responsiveness when using the plasma for measurements. Make the appropriate adjustment for your TV.

This additional lag, however, should not be ignored, and it should be taken into account when developing the game. Many people will run your game on a TV that introduces additional lag. Steps should be taken to minimize the negative effects this unavoidable lag can have.

To measure the response time (lag) the simplest tool I found has been the free Quicktime viewer. With Quicktime installed, right click on the video file from your camera, and then click on “Open with Quicktime”. You can then use the arrow keys to frame advance 1/60th at a time. (On a Mac, it works best to use Quicktime Player 7, which can be found in the Utilities folder)

For a particular response time, navigate the video to the first frame that shows the button fully pressed, and then count additional frames from there until you see the first indication of a response on screen. The number of frames you counted is your response time in 1/60ths of a second. You can convert to milliseconds by dividing by 60 and multiplying by 1000 (12/60ths is 200 ms)

Here are the frames for the PS3 system menu on the CRT TV:

In this first image (above), the thumb is descending to the “down” button on the D-Pad. It’s important to press the buttons with the thumb starting off the pad and use a rapid hitting motion, so you can see the precise frame when the button is fully pressed. Note the slight motion blur on the thumb – this is also a useful indicator of if it is still moving or not.

Now the thumb is clearly stopped and is pressing the button. Since this is the first frame where we see the button pressed, then we start counting from here, at zero (if the response were visible here, you’d have zero lag, impossible with current systems).

Now (above) we are on the first frame of actual lag, note the same line on the menu is still highlighted.

Second frame of lag, still no change

Third frame. The menu begins to move. This means that the PS3 system menu has a response time of 3/60ths of a second, or 50 ms (milliseconds). This is VERY GOOD, and is the fastest you can realistically expect from the PS3. While it’s theoretically possible to go to 2/60ths, this has severe performance problems. I don’t think 1/60th is possible on the PS3. So – all games running at 60 fps should shoot for 3/60ths as an ideal response time. This also means games running at 30 fps should shoot for 6/60ths. That’s in a perfect world, and as we shall see, actual results vary.

RESULTS FOR GTA-IV

The first thing I tried was shooting a gun in GTA-IV for the PS3 on a plasma television. This is done with the right trigger button (R2), which feels (slightly) like pulling the trigger of a real gun. In real life, there is zero delay between the trigger reaching the end of it’s travel, and the bullet leaving the gun. In GTA-IV, this delay is somewhat longer. Here’s the actual movie, so you can frame advance it yourself, and here are the relevant frames:

Note on the first frame (-1) , the finger is still moving, we start counting on the next frame (0) when the finger is fully on the button and the button is fully depressed. We then count until the first response, which comes at frame 12. This indicates a response time of 12/60ths. Since it’s measured on the plama TV, we adjust this to 10/60ths. This gives us a raw response time for GTA-IV of 166 ms (200 ms on flat panel TVs).

I repeated this test for turning and jumping, and got the same results. This is rather a long response time, and correlates with people reporting the game being sluggish and unresponsive. The delay in firing the gun after pulling the trigger is quite noticeable.

OTHER RESULTS

I tested various other games, with various results. I’ll list the results in 60ths of a second, and the results have been adjusted -2 to account for the plasma lag. All games are on the PS3 unless noted otherwise, and I’ve included the PS3 system menus and GTA for reference.

Games that run at 60 fps:

  • PS3 System menus: 3/60ths
  • Guitar Hero 3 (XBox 360): 3/60th
  • Ridge Racer 7: 4/60ths
  • Virtua Tennis: 4/60ths
  • Ninja Gaiden Sigma: 4/60ths
  • PixelJunk Racers 4/60ths

Games that run at 30 fps:

  • Genji: days of the Blade: 6/60ths
  • Tony Hawk’s Proving Ground: 8/60ths
  • Blacksite: Area51: 8/60ths
  • Halo 3 (XBox 360) : 8-10/60ths
  • EA’s “Skate”: 10/60ths
  • GTA-IV: 10/60ths
  • Harry Potter: 10-14/60ths
  • Heavenly Sword: 7-18/60ths

The first question that arises is: why are there no PS3 games with a response time of 3/60ths? The PS3 UI runs at 3/60ths response time, so it’s quite possible. Why do all the 60fps games run at 4/60ths? Still, that’s still pretty good, and those games all feel quite responsive. Guitar Hero on the XBox manages a very impressive 3/60ths – very important for that type of game, especially when it’s actually 5/60ths on many TVs.

Next, why so much variance between the 30fps games – ranging from 6/60ths to 18/60ths?

It’s worth noting that while EA’s Skate seems like it would be a bit sluggish at 10/60ths, it actually feels quite responsive, in large part due to its use of stick gestures for input, as the movement starts before the gesture is complete, yet the player mentally synchronizes it with the end of the gesture, and so it actually seems very responsive.

Then what about the games where the lag varies? Harry Potter is 10-14 which is bad enough, but Heavenly Sword is an astonishing 7-18. It takes 7/60ths to start an attack, but 18/60ths to start to turn around (See movie: mvi_4263w). Clearly something is wrong there. I would consider that a bug. It’s sad that the programmer made the effort that allowed for a 7/60th response time, but then someone else messed up down the line, making the turn take nearly a third of a second. Halo 3 is another example, with the shooting and moving being 8, but the jumping being 10.

Particularly interesting here in the 30fps category is Genji: days of the blade. This is a very similar game to Ninja Gaiden, and yet Genji runs at 30, while Ninja Gaiden runs at 60. However, the fact that Genji runs at 30 is barely noticeable, and does not detract from the game at all. In part this is because of the way the camera moves smoothly through the scene with very few rapid pans. It’s also because of the low contrast graphics and motion blur. But it’s also because Genji’s response rate is 6/60ths, very similar to Ninja Gaiden’s 4/60ths.

Ninja Gaiden is a faster paced game than Genji, with the main character jumping around rapidly, and the camera tightly following him – so the graphical benefits of 60fps are more apparent. The low response time feels good as well. However, since Genji also has a low response time, it would benefit very little from running at 60 fps. As it runs at 30fps, this gives the designers the opportunity to put more graphics, enemies and special effects on screen, and reduces the pressure on programmers and artists to constantly strive to maintain 60, which can be a difficult factor in development.

CONCLUSIONS AND SUGGESTIONS

Games that run at 60fps all seem to have a response time to 4/60ths, and while 3/60ths is possible, 4/60ths is a very good response time.

Some games running at 30fps have a response time of 8/60ths or 10/60ths (and some peak even higher). Genji shows us that a response time of 6/60ths is possible while running at 30 fps. 10/60ths can be too long, especially when combined with the processing delays in flat panel TVs which can push it up to 12/60ths. 1/5th of a second (200ms) is too long to wait for a gun to fire, and introduces annoying sluggishness when moving around or steering a car.

Some games have an inconsistent response time. Heavenly Sword varies from 7 to 18. If the system is capable of 7, then all moves should start in 7. Developers should verify ALL their response times, as other factors, such as animation, might be creating lag in specific places.

I suggest that game developers use this simple technique to measure the response time in their games, at least to verify that their assumptions are correct. If they are running at 60fps, then they should not be above 4/60ths. If they are running at 30fps, then they strive to duplicate Genji’s responsiveness of 6/60ths, and certainly not slip below 8/60ths. I suggest that they keep in mind that flat panel TVs (which are probably a majority in gamers households, and certainly in game reviewers’ households) add an additional 2 frames of lag, which makes it EVEN MORE important to keep programmed lag to a minimum.

I also suggest that game reviewers begin to use this technique to measure lag, and to include the measurement of lag in their reviews. While the subjective views of the reviewer are important and valuable, an objective measurement of response time would be a very useful additional piece of information for the person considering buying the game. Placing this information in a game review would encourage developers to produce more responsive games, which benefits everyone.

[Update Oct 2009] See this article that discusses some more refined techniques for measuring controller la:

http://www.eurogamer.net/articles/digitalfoundry-lag-factor-article

May 27th, 2008

Programming Responsiveness

[Update: This is a technical article on the causes of lag.  If you want the article on measuring lag, then click here]

Responsiveness is something that can make or break a game at first impression. This is especially true in reviews where a game with poor responsiveness will be described as being “sluggish”, “unresponsive”, “floaty” or “sloppy”. A better game might be referred to as “tight” or “responsive”. There are several factors that contribute to perceived responsiveness. This article looks at some of them from a programmer’s perspective, and offers some routes to making your game more responsive.

RESPONSE LAG

Response lag is the delay between the player triggering an event, and the player getting feedback (usually visual) that the event has occurred. If the delay is too long, then the game will feel unresponsive. Several factors contribute to the length of this response lag. If your game is unresponsive, it may well be the cumulative effects of four or five different factors. Adjusting one factor alone may not make a perceptible difference, but addressing all the factors can lead to a noticeable improvement.

Players, and sometime even designers, cannot always put into words what they feel is wrong with a particular game’s controls. Often they will try to do something that requires some synchronization, but will fail, and they won’t be able to tell you “the event happened 0.10 seconds after my input”, instead they will just tell you that the game felt “slow” or “not tight”, or “difficult”. Or they might not be able to tell you anything, and simply say the game sucked, without really understanding why it sucked. Designers and programmers need to be aware of response lag, and the negative effect it has on a game, even if test players do not directly report it as a factor.

WHY LAG HAPPENS

To understand why lag occurs, you need to understand the sequence of events that occur from the time the user presses a button, to when the results appear on screen. To understand this, we need to look at the main loop structure of the game. The main loop performs two basic tasks: logic and rendering. The “logic” portion of a main loop updates the game state (the internal representation of the game objects and environment), while the “rendering” portion of the loop creates a frame that is displayed on the television. At some point in the main loop (usually at the start) we also get input from the user. This is sometimes considered a third task in the main loop, but it’s also commonly a part of the logic task. I’ve kept it separate here because it’s important to see in what order things happen

LISTING 1 – THE SIMPLEST MAIN LOOP

while (1) {
  Input();
  Logic();
  Rendering();
}

There are several ways a main loop can be structured. The simplest is shown in listing 1, where we simply alternate calling the logic and the rendering code. It is assumed there is some frame synchronization in the call to Rendering(), and that we are running at a fixed frame rate, usually 60 or 30 frames per second for an NTSC console game.

The main loop here is also only showing half the story. The call to Rendering() is doing the CPU side of the rendering task, which is iterating over the environment and the objects, culling, animating, sorting, setting up transforms, and building a display list for the GPU to execute. The actual GPU rendering is performed after the CPU rendering, and usually is asynchronous, so while the main loop is processing the next frame, the GPU is still rendering the previous frame.

So where does the lag come in? Look at the sequence of events that occur from the user pressing a button to there being some feedback for pressing that button. A the highest level, the user presses a button, the game logic reads that button press and updates the game state, the CPU render function sets up a frame with this new game state, then the GPU renders it, and finally this new frame is displayed on the television.

Figure 1 shows this in a more graphical manner. Sometime in frame 1, the player presses a button to fire his gun. Since the input processing has already been done for that frame, this input is read in frame 2. Frame 2 updates the logic state based on this button press (a shot is fired). Also in frame 2, the CPU side of rendering is performed with this new logic state. Then on frame 3, the GPU performs the actual rendering of this new logic state. Finally at the start of frame 4, the newly rendered frame is presented to the user by flipping the frame buffers.

So how long is the lag? It depends on how long a frame is (where a “frame” here is a complete iteration of the main loop). It takes up to three frames for the users input to be translated into visual feedback. So if we are running at 30 frames per second, then the lag is 3/30th or 1/10th of a second. If we are running at 60 frames per second, then the lag will be 3/60th or 1/20th of a second.

This illustrates a common misconception about the difference between 60 fps and 30fps games. Since the difference between these two frame rates is just 1/60th of a second, people assume that the difference in responsiveness will also be 1/60th. In actual fact, going from 60 to 30 does not just add a vsync to your lag, it acts as a multiplier, doubling the length of the process pipeline that is responsible for lag. In our ideal example in figure 1, it adds 3/60ths of a second, not 1/60th. If the event pipeline is longer, which it quite possible can be, then it can add even more.

Figure 1 actually illustrates the best possible sequence of events. The button press event is translated into visual feedback via the shortest path possible. We know this because we can clearly see the sequence of events. As a programmer, being familiar with the order in which things happen is a vital part of understanding why thing act the way they do in the game. It quite easy to introduce additional frames of lag (meaning an extra 1/60th or 1/30th of a second delay), by not paying careful attention to the order in which things happen.

As a simple example, consider what would happen if we switched the order of the Logic() and Rendering() calls in our main loop. Look at frame 2 of figure 1, although the input is being read at the start of frame 2, and affects the logic during frame 2, the rendering is performed first, and so is using the logic state from frame 1. This introduces an extra frame of lag. While this is a novice level mistake, it is still something you need to make absolutely sure is not happening.

Extra frames of lag can be introduced in a more subtle manner as a result of the order of operations within the game logic. In out example we are firing a gun. Now perhaps our engine is set up to increment the position of the objects in the world using a physics engine, and then handle events that are raised due to this update (such as collision events). So in this situation the sequence of input/logic looks like listing 2.

LISTING 2 – PHYSICS UPDATE IS FOLLOWED BY EVENT HANDLING

void Logic() {
  HandleInput();
  UpdatePhysics();
  HandleEvents();
}

Now event handling via messages is a very nice way of decoupling systems. So a programmer might decide to use it for the player control events. To fire a gun, the HandleInput() function will fire an event telling the gun to fire. The HandleEvents() function will take this event and cause the gun to actually fire. But because the physics update has already happened for this frame, the effect on the world state will not be incorporated until the next frame, hence introducing an extra frame of lag.

MORE LAG CAUSES

The lag can be extended even further by lower level action ordering. Consider a “jump” – the feedback comes when the character actually moves. To make something move in a game, you can either set the velocity directly, or you apply a force to it, either acceleration or, more likely, a momentary impulse. The problem arises here if in your physics engine then positions are updated before the velocity change is applied (a common situation in many introductory game programming tutorials). In this case, although the velocity of the jumping object will be set on the same frame as the input event is handled, the object will not actually begin to move until the next time around the loop, on the next game frame, and so introduces an additional frame of lag.

Remember these are cumulative problems that can be difficult to discern in isolation, but the combined effect can make your game controls turn to mush. Suppose you had made all three mistakes listed above: you do rendering before logic, you handle logic events after advancing the physics state, and you update position before velocity. That’s three additional full iteration of the main loop, in addition to the three you already have built in. Six frames of lag between the player pressing a button, and seeing the result on screen. At 60 frames per second that’s 1/10th of a second, which is bad enough, but if the game is running at 30 frames per second, then the lag is DOUBLED to an unbearable 1/5th of a second, or 200ms.

Other factors can contribute to lag, compounding even further the effects above. Movement can be driven by animations, with velocity changes built into specific time points in an animation. The obvious problem here is if the animator places the jump velocity impulse a fraction of a second into the animation, to better match the visuals. It might look better, but it feels bad. That’s easily dealt with by making sure the velocity impulse is on the first frame of animation when immediate feedback is needed. But then the question is: how does triggering an animation translate into actual movement? It’s quite likely that animation updating is handled by the Render() function, and so any events triggered by the animation will not be handled until the time around the loop, which adds another frame. In addition, triggering an animation might not make it advance a frame until the next frame, delaying the event firing for a frame. Our lag could potentially be increased from six to eight frames, which would be quite unplayable, even at 60 frames per second.

That’s not the end of it either. There are many ways in which extra frames of lag can be introduced. You might be pipelining your physics on a separate thread (or a physics processing unit). You might be using triple buffering to smooth your frame rate. You could be using abstract events that take a couple of passes through the system to resolve into real events. You could be using a scripting language that adds an additional frame in the way it waits for an event. It’s quite possible to make you game logic incredibly flexible by abstracting various concepts of time and events, and yet while doing this you loose sight of exactly what is going on under the hood, making it far easier for additional frames of delay to creep in.

RESPONSIVENESS IS NOT REACTION TIME

On of the great misconceptions regarding responsiveness is that it is somehow connected to human reaction time. Humans cannot physically react to a visual stimulus and then move their fingers in less than one tenth of a second, and peak reaction times vary from 0.15 seconds to 0.30 seconds, depending on how “twitchy” the gamer is. Figures such as these are often brought up in discussion of game responsiveness, but the connection is specious.

It’s not how fast you react to the game that is the issue; it’s how fast the game reacts to you. The issue is not one of reaction times, but of synchronization. In games such as Guitar Hero there are things coming towards you, and you have to hit the correct button at a very precise point in time, when the target object is within a particular region. The player is anticipating the future event, and there is no reaction involved at all. The problems of lack of responsiveness occur when the game does not react fast enough to the player, and the target object has moved beyond the target region by the time the event occurs. The player presses the button at the correct point, and they do not expect the object to move even a few more pixels before it explodes. But since objects generally move at least a few pixels per frame, having a few frames of lag can allow the object to drift past its target.

Many action games are based around this anticipation and button pressing. In a skateboarding game you want to jump just before you hit the end of a rail. In a shooter, you want to shoot the instant someone moves in front of your gun. Again this is not reaction time, you will usually have seen the target at least half a second before you shoot it, probably more, and will be either moving the gun, or waiting for the target to move in front of the gun.

Because of the somewhat unintuitive nature of these problems, in order to create a responsive game, it is important that you fully understand the issues involved. The most important thing is to be able to clearly describe the frame by frame path though your logic and rendering that the action a button being pressed will take in order to create visual feedback. Once you have this, you can optimize it as close to the optimal pathway as possible.

|