Love is patient, love is kind.  It does not envy, it does not boast, it is not proud.  It is not rude, it is not self-seeking, it is not easily angered, it keeps no record of wrongs.  Love does not delight in evil but rejoices with the truth.  It always protects, always trusts, always hopes, always perseveres.

1 Corinthians 13

Love is easy.  Middleware is hard.

We used middleware on Fracture for physics, tree rendering, audio, animation, facial animation, network transport, and various other systems.  Now that the game is finished, we’re taking stock of the lessons that we learned and deciding what packages we want to keep and what we want to replace as we move forward.

Middleware offers two main benefits, each of which is balanced by an associated cost:

1.  Middleware provides you with more code than you could write yourself for a fraction of what it would cost you to try.  No matter how clever you are, that’s how the economics of the situation work.  The costs of developing a piece of software are largely fixed.  But once written, any piece of software can be sold over and over again.  Because they can spread their costs over a large number of customers, middleware vendors can afford to keep larger, more experienced teams working on a given piece of functionality than independent game developers can.  During Fracture, we watched a couple of pieces of immature middleware spring up after we’d written similar tools ourselves.  We watched those pieces of middleware grow and surpass our tools, because we couldn’t afford to devote the resources to our tools that the middleware vendors devoted to theirs.

The corresponding cost is that none of the functionality you get is exactly what you would have written yourself, and much of it will be entirely useless to you.  Middleware is written to support the most common cases.  If you have specific needs that differ from the norm, you may be better off implementing a solution yourself.

2.  Middleware offers structure.  Middleware draws a line between the things that you have to worry about and the things you don’t.  As long as it’s reasonably well documented and stable, you don’t need to waste mental bandwidth worrying about the things that go on underneath your middleware’s public API.  As games grow ever-larger and more complex it’s become incredibly valuable to be able to draw a line and say, The stuff on the other side of that line isn’t my responsibility, and I don’t have to worry about it.

The associated cost is that you can’t change what’s on the other side of that line.  If you’re going to use middleware, you have to be willing to accept a certain amount of inflexibility in dealing with the problems that the middleware solves.  You have to be willing to shape your own technology to suit the third-party libraries that you’re buying.  Trying to do otherwise is a recipe for misery.

Given those benefits and costs, it makes sense to use middleware wherever you can–as long as you don’t try to license technology in the areas where you want your game to be unique.  Every game has certain unique selling propositions:  things that make it distinct from every other game on the market.  Likewise, every game has certain characteristics that it shares with many others.  When it comes to the latter, you should buy off-the-shelf technology, accept that technology’s inflexibility, and modify your game design to suit it.  When it comes to the former, you should write the code in-house.  As Joel Spolsky says, don’t outsource your core competency.

So now that you’ve figured out which game functions you should be implementing through middleware, how do you decide which of the scads of available middleware packages is best for you?  There are a number of issues to keep in mind.

Good middleware lets you hook your own memory allocator.  If your approach to memory is to partition the entire free space up front and minimize runtime allocations, then you’ll want the ability to reserve a block for your middleware and allocate out of that.  Even if you allow dynamic allocations at any time, you’ll want to track how much memory is used by each system and instrument allocations to detect memory leaks.  Any piece of middleware that goes behind your back and allocates memory directly just isn’t worth buying.

Good middleware lets you hook your own I/O functions.  Most games store resources in package files like the old Doom WAD files.  Middleware that doesn’t let you hook file I/O doesn’t let you put its resources in packfiles.  Many modern games stream resources off DVD as the player progresses through a level.  If you can’t control middleware I/O operations, then you can’t sort file accesses to minimize DVD seeks.  You’ll waste read bandwidth and your game will be subject to unpredictable hitches.  Again, any piece of middleware that does file I/O directly just isn’t worth buying.

Good middleware has extensible functionality.  No middleware package will do everything you want out of the box.  But you shouldn’t have to modify any piece of middleware to make it do what you need.  Good middleware offers abstract interfaces that you can implement and callbacks that you can hook where you need to do something unique to your game.  An animation package may let you implement custom animation controllers.  A physics package may let you write your own collision primitives.  Your objects should be first-class citizens of your middleware’s world.

Good middleware avoids symbol conflicts.  Beware of middleware that uses the Standard Template Library carelessly.  It’ll work fine until you try to switch to STLPort or upgrade to a new development environment, and then you’ll suddenly find that your engine has multiple conflicting definitions of std::string and other common classes.  To avoid symbol clashes, every class in a middleware library should start with a custom prefix or be scoped inside a library namespace.  And if a piece of middleware is going to use the STL, it should do so carefully, making sure that every STL class instantiated uses a custom allocator.  That allows you to hook your own memory allocator and avoids symbol conflicts.

Good middleware is explicit about its thread safety.  We live in an ever-more-multithreaded world, but one where most game engines are still bound by main thread operations most of the time.  For best performance, you want to offload any operations you can onto other threads.  To do that with a piece of middleware, you need to know which operations can be performed concurrently and which have to happen sequentially.  Ideally a piece of middleware will let you create resources asynchronously so you can construct objects in a loader thread before handing them off to the game.

Good middleware fits into your data pipeline.  Most companies export data in an inefficient platform-independent format, then cook it into optimized platform-specific formats and build package files out of those as part of a resource build.  Any piece of middleware should allow content creators to export their assets in a platform-independent format.  Ideally, that format should still be directly loadable.  The build process should be able to generate platform-specific versions of those assets using command-line tools.

Good middleware is stable.  One of the main benefits of middleware is that it frees your mind to focus on more critical things.  In this respect, middleware is like your compiler:  It frees you from having to think about low-level implementation details–but only as long as you trust the compiler!  A buggy piece of middleware is a double curse, because instead of freeing your attention from a piece of functionality, it forces you to focus your attention there, on code that was written by a stranger and that no one in your company understands.  Worse still, if you’re forced to make bug fixes yourself, then you need to carry them forward with each new code drop you get of your middleware libraries.  You shouldn’t need to concern yourself with the implementation details of your middleware.  Middleware is only a benefit to the extent that its API remains inviolate.

Good middleware gives you source.  Despite the previous point, having access to the source for any middleware package is a must.  Sometimes you’ll suspect that there’s a bug in the middleware.  Sometimes you need to see how a particular input led to a particular result before you can understand why the input was wrong.  Sometimes you’ll have to fix a bug no matter how good the middleware is.  And frequently you’ll need to recompile to handle a new platform SDK release or to link with some esoteric build configuration.

There are other questions to ask about any piece of middleware:  How much memory does it use?  How much CPU time does it require?  What’s the upgrade path for your current code and data?  How does it interact with your other middleware?  How good is the vendor’s support?  How much does it cost?

But those questions have vaguer answers.  Acceptable performance or cost will vary depending on the nature of your product.  A cell phone game probably can’t afford to license the Unreal Engine–and probably couldn’t fit it in available memory if it did!  Data upgrade paths are less of a concern if you’re writing a new engine from scratch than if you’re making the umpteenth version of an annual football game.

The rules that don’t vary are:  You should buy middleware wherever you can do so without outsourcing your core competency.  And to be worth buying, any piece of middleware should behave itself with respect to resource management and concurrency.

Many systems naturally lend themselves to a design that gathers the high-level functionality of the system behind a single interface:  A FileManager, for example, might expose functions for building a file database from packfiles and loose files.  An AudioManager might expose functions for loading and playing sound cues.  A SceneManager might expose functions for loading, moving and rendering models.

Once upon a time, these objects would all have been global variables.  There are a number of problems with using global objects as managers, though, the most critical of which is the uncertainty of static initialization and destruction order.  Your managers start out in a pristine garden of Eden, free of all knowledge of good, evil, and (most importantly) one another.  But as your game gets more and more complicated, your managers are going to develop more dependencies on one another.  To add streaming of sounds to AudioManager, for example, you may need to make it reference FileManager.  To stream asynchronously, you may also need to reference AsyncJobManager.  But at construction time, you can’t be sure that any of those managers exist yet.

This prompts clever people who’ve read Design Patterns to think, Aha!  I’ll make all my managers Meyers singletons!  And they go and write something like

FileManager& GetFileManager()
{ 
    static FileManager theFileManager; 
    return theFileManager;
}

This function will create a FileManager object the first time it’s called and register it for destruction at program exit.  There are still a couple of problems with that, though:

• First, like all function-local static variables in C++98, theFileManager is inherently unthreadsafe.  If two threads happen to call GetFileManager simultaneously, and theFileManager hasn’t been constructed yet, they’ll both try to construct it.  Wackiness ensues.  This quibble is due to be fixed in C++0x.
• Imagine what happens if the constructor for AudioManager calls GetFileManager and the constructor for FileManager calls GetAudioManager.  To quote chapter and verse of the ISO C++ Standard, “If control re-enters the declaration (recursively) while the [static] object is being initialized, the behavior is undefined.” (ISO Standard 6.7.4) The compiler can do whatever it wants, and whatever it does is unlikely to be what you want.
• Although Meyers singletons give you a JIT safe order of construction, they make no promises about their order of destruction.  If AudioManager is destroyed at program exit time, but FileManager wants to call one last function in AudioManager from its destructor, then again you have undefined behavior.

We dealt with those issues in MechAssault 2 by eschewing automatic singletons in favor of explicitly constructed and destroyed singletons.  Every manager-type object had an interface that looked something like 

class FileManager
{
public:  
    static bool CreateInstance();
    static void DestroyInstance();
    static FileManager* GetInstance();
};

This works better than automatic singletons.  It worked well enough that we were able to ship a couple of quite successful games with this approach.  We wrote a high-level function in the game that constructed all the managers in some order and another high-level function that shut them all down again.  This was thread-safe, explicitly ordered and deterministic in its order of destruction.

But as we added more and more library managers, cracks started to show.  The main problem was that although we constructed managers in an explicit order, dependencies between managers were still implicit.  Suppose, for example, that AudioManager is created before FileManager, but that you’re tasked with scanning the filesystem for audio files at start-up time.  That makes AudioManager dependent on FileManager for the first time.  Now AudioManager needs to be created after FileManager.

Changing the order in which managers were constructed was always fraught with peril.  Finding a new working order was a time-consuming process because ordering failures weren’t apparent at compile time.  To catch them, you needed to actually run the game and see what code would crash dereferencing null pointers.  And once a new ordering was found, it needed to be propagated to every application that used the reordered managers.  Day 1 has always placed a strong emphasis on its content creation tools, and most of those tools link with some subset of the game’s core libraries, so a change might need to be duplicated in six or eight tools, all of which would always compile–but possibly fail at runtime.

With Despair, one of our governing principles has been to make implicit relationships explicit.  As applied to system manager construction, that meant that managers would no longer have static GetInstance functions that could be called whether the manager existed or not.  Instead, each manager takes pointers to other managers as constructor parameters.  As long as you don’t try to pass a manager’s constructor a null pointer (which will assert), any order-of-initialization errors will be compile-time failures.  To make sure that managers are also destroyed in a valid order, we use ref-counted smart pointers to refer from one manager to another.  As long any manager exists, the managers that it requires will also exist.

Our current code looks something like  

class IAudioManager
{
public:
   IAudioManagerPtr Create(const IFileManagerPtr& fileMgr);
};

class AudioManager
{
public:
   explicit AudioManager(const IFileManagerPtr& fileMgr);
};


One problem remains.  Although AudioManager exposes the public interface of the audio library to the rest of the application, there are also private support and utility functions that the audio system should only expose to its own resources.  Furthermore, many of these utility functions will probably need access to lower-level library managers.

This could be solved by having global utility functions that take manager references as parameters and by making every audio resource hold onto pointers to every lower-level manager that it might use.  But that would bloat our objects and add reference-counting overhead.  A more efficient and better-encapsulated solution was to give each library a context object.  As Allan Kelly describes in his paper on the Encapsulated Context pattern:  

A system contains data that must be generally available to divergent parts of the system but we wish to avoid using long parameter lists to functions or globaldata.  Therefore, we place the necessary data in a Context Object and pass this object from function to function.    

In our case, that meant wrapping up all smart pointers to lower-level libraries in a single library context, which would in turn be owned by the library manager and (for maximum safety) by all other objects in the library that need to access a lower-level manager.  Over time, other private library shared state has crept into our library contexts as well.  To maintain a strict ordering, the library manager generally references all the objects that it creates, and they reference the library context, but know nothing about the library manager.

This architecture has generally worked well for us, and has been entirely successful in avoiding the manager order-of-creation and order-of-shutdown issues that plagued earlier games we worked on.  It does have definite costs, however:

•  A lot of typing.  Library managers know about library resource objects which know about library contexts.  But frequently there’s some functionality that you’d like to be accessible both to objects outside the library and objects inside the library.  In this architecture, there’s nowhere to put that functionality except in the library context, with pass-through functions in the library manager.  Since the library manager is usually hidden behind an abstract interface, you can end up adding function declarations to three headers and implementations to two cpp files before you’re done.
• Even more typing.  All the lower-level managers held by a library also get passed through a creation function to the library manager’s constructor to the library context’s constructor.  That was a mild annoyance when the engine was small and managers were few, but Despair now comprises over fifty libraries.  Most of those libraries have managers, and there are a handful of high-level library managers that take almost all lower-level library managers as constructor parameters.  Managers that take forty or fifty smart pointers as parameters are hard to create and slow to compile.
• Reference counting woes.  In principle, every object should strongly reference its library context to maintain a strict hierarchy of ownership and to make sure that nothing is left accessing freed memory at program shutdown time.  In practice, though, this doesn’t work well when objects can be created in multiple threads.  Without interlocked operations, your reference counts will get out of sync and eventually your program will probably crash.  But with interlocked operations, adding references to library contexts becomes much more expensive, and can become a significant part of your object creation costs.  In practice, we’ve ended up using raw pointers to contexts in several libraries where the extra safety wasn’t worth the additional reference cost.

So managers and contexts are far from perfect.  They’re just the best way we’ve found so far to stay safe in an ever-more complex virtual world.

Nonetheless, every game engine we’d worked on before Despair had a similar deep inheritance hierarchy of the sort that was in vogue in the mid-nineties:  a player class might inherit from some kind of combatant class, which would inherit from a mover class, which would inherit from a physical object class, which would inherit from a base game object class.

This architecture has a lot of shortcomings.  Let me enumerate a few of them:

First, it’s inflexible.  If you want to create a new AI enemy that has some of the capabilities of enemy A and some of the capabilities of enemy B, that’s not a task that fits naturally into a hierarchical object classification.  Ideally, you’d like the designers to be able to create a new enemy type without involving you, the programmer, at all.  But with a deep object hierarchy, you have to get involved and you have to try to pick the best implementation from several bad options:  to have your new enemy class inherit from one object and cut-and-paste the functions you need from the other; to not inherit from either, and to cut-and-paste the functions you need from both; or to tiptoe down the treacherous slippery slope of multiple inheritance and hope that it doesn’t lead to a diamond of death.

Second, a handful of classes in your hierarchy tend to grow without bound over a game’s development.  If the player class is part of the object hierarchy, then you can expect this class to include input and control systems, custom animation controls, pickup and inventory systems, targeting assistance, network synchronization–plus any special systems required by the particular game that you’re making.  One previous game that we worked on features a 13,000 line player class implementation, and the player class inherited from a 12,000 physical object class.  It’s hard to find anything in files that size, and they’re frequent spots for merge conflicts since everyone’s trying to add new stuff to them all the time.

Third, deep inheritance is poor physical structure.  If class A inherits from class B which inherits from class C, then the header file for A–A.h–has to #include B.h and C.h.  As your hierarchy gets deeper, you’ll find that all of your leaf classes have to include four or five extra headers for their base classes at different levels.  For most modern games, as much compile time is spent opening and reading header files as is spent actually compiling code.  The more loosely your code is coupled, the faster you can compile.  (See Lakos for more details.)

Therefore we resolved to make Despair as component-based as we could, and to keep our inheritance hierarchies as flat as possible.  A game object in Despair, for example, is basically a thin wrapper around a UID and a standard vector of game components.  Components can be queried for type information and dynamically cast.  The game object provides only lifetime management and identifier scoping.  It knows nothing about component implementations.  It contains no traditional game object state like position, bounds, or visual representation.

This approach has informed other systems as well.  Our scene object implementation is similar to the game object implementation, with a single object representing each model that provides lifetime management for a vector of scene nodes.  Scene nodes manage their own hierarchies for render state or skeletal transforms.

Another family of systems is built on a flow-graph library for visual editing.  Game logic, animation systems, and materials can all be built by non-programmers wiring together graph components in the appropriate tools.

Using composition instead of inheritance has worked very well for us.  Our primary concern when we set out in this new direction was that we’d end up with something that had horrible runtime performance.  With Fracture almost complete, though, there’s no evidence that our performance is worse than it would have been with a deep inheritance hierarchy.  If anything, I’m inclined to suspect that it’s better, since well-encapsulated components have better cache locality than large objects and since the fact that we only update dirty components each frame means that we can decide what does and doesn’t need to be updated at a finer granularity.

If I were starting over again, the only change I’d make with respect to object composition is to make scene objects more opaque and less like game objects.  Scene objects have a different problem to solve.  We have several hundred game components now, with more going in all the time, and the flexibility of having a thin game object interface that allows querying components for type has paid big dividends.  I think that our game object system is close to ideal for iterating rapidly on gameplay.  Scene objects are a different kind of problem, though.  We haven’t added any new scene node types since the scene library was written, and all scene nodes are implemented in the scene library instead of in higher-level code.  At the same time, it would be nice to be able to experiment with different optimizations of updating skeletal hierarchies without breaking higher-level code.  All of this argues that following the Law of Demeter and hiding scene object implementation details would have been appropriate for scene objects even if it wasn’t appropriate in the rapid-prototyping environment of game objects.

Beyond the perfect abstract world of software architecture, component based design also created a couple of surprises in the messier world of development process and human interaction.  One lesson of working with a composition-based engine is that the learning curve for new programmers is steeper.  For programmers who are used to being able to step through a few big nested functions and see the whole game, it can be disorienting to step through the game and discover that there’s just no there there.  For example, Despair contains over 500 classes with the word “component” in their names and 100 classes with the word “object” in their names.  Our games aren’t defined by C++ objects, they’re defined by relationships between them.  To understand those relationships, you need good documentation and communication more than ever.

Another composition lesson is that component-based design takes a lot of the complexity that used to exist in code and pushes it into data.  Designers aren’t used to designing objects and constructing inheritance hierarchies.  Working with components requires new processes and good people.  Cross-pollination is important.  Your programmers need to work building objects in your tools, as well as writing code for components, and they need to work hand-in-hand with good technical designers who can provide tool feedback and build on the object primitives available to them.  Like a game, your team isn’t defined by its individual components but by the relationships between them.

The engine that we used at Cyan, for example, was a pure scene graph.  Every part of the game was represented by one or more nodes in a hierarchy.  But the hierarchy represented logical relationships in some places and kinematic relationships in others.  Throughout the graph, ownership was conflated with update order:  children would be deleted when their parents were deleted and parents always updated before their children.  Kinematic attachment was performed by pruning and grafting trees in the graph, which had the effect of tying the lifetimes of attached objects to the lifetimes of their parents.

The engine used for MechAssault 2 was similar, except that it had separate game object and scene hierarchies.  In both hierarchies, graph relationships defined ownership and update order:  children would be deleted when their parents were deleted and parents were always updated before their children.  In the game object graph, hierarchy also defined logical relationships that varied based on the objects involved.  Triggers, for example, would treat certain of their child objects as inputs to be checked and others as objects to be acted on.  In principle, the entire game object hierarchy updated, then the entire scene hierarchy updated.  In practice, we needed to know the final animated positions of some objects so that we knew where to create effects, so some trees in the scene hierarchy got explicitly updated during the game object update.

All of this was very confusing.  It also meant that changes to the hierarchy motivated by one concern tended to have unforeseen side-effects in other areas.  And it necessitated odd hacks like the MechAssault’s explicit updating of sub-trees in the scene graph. In developing Despair, therefore, we decided to have one data-structure per purpose.  As mentioned previously, we have a well-defined ownership hierarchy among our objects.  For example,

• The streaming manager owns worlds
• Worlds own game objects
• Game objects own game components
• Game components own scene objects (and audio resources, and HUD elements, etc…)
• Scene objects own mesh nodes (and transform nodes, and light nodes, etc…)
• Mesh nodes own vertex groups and index lists

In some cases ownership is shared and in some cases it isn’t, but each object has its level of the hierarchy.  But in addition to the hierarchy that we use for ownership, we also have parallel data structures for

• the object registry used for object look-up
• the dependency graph used for object update
• the physical representation of the world (stored in Havok’s internal hierarchy)
• the renderable representation of the world (stored in our own sphere tree of renderable objects)
• etc.

This has been a big win in terms of clarity and flexibility.  It’s clear when looking at a component what its lifetime will be.  It’s easy to make a component update later in the frame, or not update at all, without having to worry that your change will affect when the object is deleted or whether it’s rendered.  Scene objects and game components are part of the same dependency hierarchy.  Therefore, the component that creates a fire effect can make itself dependent on the scene object for the player’s weapon.  That way the fire effect won’t be created until the weapon has determined its final position for the frame, and the fire effect will be created at the correct location.

The performance implications of orthogonal views are unclear.  One of the frustrations of building a game engine is that you never have the leisure of changing one architectural variable while keeping all others constant.  That means that you can never be sure you’ve made the best decision.  You can only be sure–sometimes–when you’ve made a very bad one.Having orthogonal views of our data was clearly not a decision worthy of regret.  But it’s difficult to be certain whether it’s a net win in terms of performance.  On the negative side, all these different hierarchies are scattered at widely different locations in memory, and on modern hardware memory locality drives performance more than the number of instructions executed does.  But on the other hand, having independent views of the data means that each only needs to process what matters.  Our update pass only updates objects that are actually changing.  Our render pass doesn’t waste any overhead on objects that aren’t renderable.  And when those systems operate, they perform similar operations on sorted lists of similar objects, which is the sort of thing that modern hardware has gotten very good at.  So I think this idiom is a performance win, but the only hard data I have to back that up is that we’re spending less time waiting on cache misses than we were in MechAssault 2.  Given the other differences between Despair and the MechAssault 2 engine and between the Xbox 360 and the original Xbox, that evidence is suggestive but not conclusive.

The one area in which orthogonal views have a clear cost is memory.  As objects add themselves to all our different systems, they may have representations in the registry, in the dependency graph, the scene hierarchy, the physics hierarchy, the render hierarchy, and probably one or two other systems that I’m forgetting.  I’d estimate that this overhead adds up to somewhere between two and three megabytes for a typical level in Fracture.  On a console with 512 megabytes of memory that cost seems justifiable, but it would have been less practical on an Xbox or PlayStation 2 game.

We wanted to improve stability and leak less memory in Despair, and this led us to consider exception safety.  Exception safety is the idea that a program should always remain in a valid state in the presence of exceptions.  Operations should succeed or fail, but if they fail, they should have no side effects and program data should remain unchanged.  Memory leaks frequently occur when exception unsafe code encounters an exceptional situation.  A function may throw an exception to signal bad data, for example, and thereby skip clean-up of temporary objects done at the end of the function.

Despair doesn’t use exceptions, because they introduce significant overhead, but we do use error codes and FAIL_RETURN macros that may early exit from a function.  These introduce similar problems.  Therefore we try to be exception-safe wherever possible without hurting performance.  In practice, performance concerns generally keep us from using the PIMPL idiom to implement atomic transactions.  But we do use the Resource Acquisition Is Initialization idiom extensively.  Our XML writer uses scoped EnterElement/LeaveElement objects.  Our critical sections use scoped locks.  And we use smart pointers extensively.

The Despair Engine contains about ten smart pointer types, depending on how broadly you define a smart pointer.  I’ve had very smart, experienced engineers roll their eyes and say, “One smart pointer type is too many!” when I tell them that, but in actual practice they’ve worked so well for us that I couldn’t imagine working again on an engine that didn’t use smart pointers.

There are some things that must be borne in mind when using smart pointers:

• Smart pointers don’t remove any of the decisions about object ownership that have to be made in designing an engine.  You still need to decide what objects are shared, if any, and you need to define a strict ordering of which objects own which other objects.  All that smart pointers do is help document that ordering (you know that the object with a scoped_ptr owns the one without, even if there are pointers going both ways) and help enforce that ownership across the space of possible execution paths.
• Smart pointers are not the same as raw pointers.  They may use pointer syntax, but it’s impossible to use them efficiently without understanding what they are and understanding when they’re going to free memory or call AddRef/Release functions.  Smart pointers passed as function parameters should almost always be passed by const reference.

Our smart pointer menagerie includes

• a pointer types for internally-refcounted objects
• a pointer type for externally-refcounted objects (rarely used)
• scoped pointers (a std::auto_ptr replacement that’s sane)
• scoped arrays
• resource pointers that start NULL and are filled in as resources load asynchronously
• intrusive weak pointers
• and a few other specific pointers for managing internally-refcounted objects in third-party APIs

Programmers are encouraged to establish strong references at load time and use them at runtime.  Because we don’t generally create or pass around smart pointers during a frame, the runtime cost of smart pointers in Despair is minimal.

I credit our use of smart pointers, our rules for ownership, and our use of the RAII idiom for giving us an engine more stable than any I’ve ever worked on before, even as the number of engineers writing code in it has grown from two to thirty.  A new memory leak shows up about every month or two, and is almost always the result of someone’s not using smart pointers and then forgetting to call delete in all code paths.  In our engine, if you find yourself having to type out the keyword “delete”, then you’re probably doing something wrong.

Another, less frequent, error that sometimes occurs is for someone to introduce a cyclic reference that causes approximately every object in the game to leak.  In Despair, this happens about once every six months.  This is not a consequence of our use of smart pointers.  Most of the other game engines that I’ve worked on had manually-refcounted objects, and had similar issues on occasion.  Despair’s smart-pointer classes do help when we’re trying to debug this situation, though.  A debug macro enables capturing the callstack on each AddRef and associating each callstack with a particular smart pointer.  We can thus report the callstack of all leaked AddRef calls for each refcounted object.  Like I said, we don’t need this capability very often, but it’s nice to have.

The name started as a joke.  Adrian and I were out to lunch with our lead one day.  Somebody made a crack about naming the engine “Despair.”  One thing led to another, and by the end of the meal we’d plotted out a whole suite of despair-themed content creation tools, most of which never got made.  It was 2004.  We were starting from scratch on core technology for big-budget AAA games running on consoles that didn’t even exist yet.  We figured that “Despair” would work one way or the other, either as an imperative to our competitors if we succeeded or as a sadly accurate description of our own feelings if we failed.

There are two broad opposing ideologies when it comes to engine design for games.

Call the first the Low-Level Faction.  This school of design traces its roots to embedded devices and to early consoles like the Dreamcast and the first PlayStation.  This school is characterized by a preference for coding close to the metal, tends to produce games rather than game engines, and eschews abstraction in favor of raw performance.  Many of the games produced by this school of thought are raw C rather than C++, or if they are C++, they tend to minimize use of C++ features like polymorphism, virtual functions and templates.  When Insomniac’s Mike Action says at GDC that “Software is not a platform!  Hardware is a platform!” he’s voicing this faction’s frustrations.

Call the other group the Abstraction Faction.  This school of game engine development grew out of PC game development, where audio, graphics and I/O devices are hidden behind a driver layer.  Programming straight to the hardware simply isn’t an option.  The programmers of this school learned their trade on a platform with virtual memory and highly variable capabilities, so they learned to use levels of indirection to abstract away platform differences.  When the Epic guys say they’d sacrifice 10% of their runtime performance for a 50% improvement in content creation time, they’re justifying their higher level of abstraction.

Both ideologies of engine development have produced great games.  Halo, for example, is a straight-C game of the Low-Level Faction.  Gears of War is an example of a game with a lot of abstraction that uses lots of script and the advanced features of C++.

The lesson that I’ve learned in a decade of software development is that the specific decisions you make are less important than that decisions are made at all.  Software requires structure.  Every decision has tradeoffs, but once made, even seemingly-arbitrary decisions have to be enforced, or entropy will consume your codebase with a vengeance.  Every-man-for-himself design will get you a million lines of code, any one of which may reference any other, and any one of which may have undefined side effects.

The Despair Engine is an unapologetic representative of the Abstraction Faction of game engine design.  This is not because I think that’s the only way to create games.  It’s because that’s the way that I know how to create games.  We use the advanced features of C++ very aggressively.  If our history of uncovering compiler errors is any indication, the Despair Engine is unique among game engines in the degree to which it pushes the infrequently-explored edges of the language.  We use STL containers and algorithms.  We use the Boost library of C++ template extensions.  We use template metaprogramming to create our own embedded domain-specific reflection language.  And we use abstraction and polymorphism throughout our code.

Over the coming weeks, I’d like to discuss some of the Despair’s more unusual idioms and axioms.  For each, I’ll describe the problem that our approach was intended to solve, and I’ll talk about the continuing challenges of the solution that we implemented.