Preventing Complexity in Game Programming

Preventing Complexity

Yaser Zhian
November 2010
Fanafzar Game Studios

What this is about
• The things that you should do at the beginning of a game
(or similar) project to simplify debugging, adding
features, optimization, build and release.
• We are talking about the technical and code aspect of
this problem, but some of the ideas are generalizable to
the rest of the team.
• I can only give you some tips, not a streamlined process.
That's your job.
• Seasoned programmers and developers will definitely
know what I'm talking about and will agree with most of
this. Less-experienced programmers have to trust me (to
some extent.)

Process, Process, Process (1/3)
• Do {Conceptualize -> Build -> Import ->
Integrate -> Test } while (0 == 0);
• Iteration time is the most important and under-
estimated factor in game development process.
• Ideally, you should never have to exit the game
to change anything (code, assets, configuration,
scripts,...)
• More realistically, you should never have to exit
the game to change anything except for code.

• Data-driven design and code is king!
• Keep engineers out of the loop. Seriously!
• Code reviews are good things, even
informal ones.
• Agile philosophy is fantastic for small
game developers.

• Automate as much as you can, as early as
possible, and keep adapting. Tools, tools,
tools!
• Transparency and discoverability
• Source control, asset control, versioning,
branches, tags and issue tracking are not
just fancy words.

Naming and Directory Structure
• A name is a unifying concept. Helps
develop common views and more efficient
and less error-prone communication.
• ??? (No one knows what you need!)
• When choosing a name, consider how that
name looks to these entities:
– Yourself, in 6 months
– Other people
– Scripts and tools

Build System (1/4)
• Quick, configurable, dependable, minimal.
• Think about header and source layout, very
carefully and very deliberately.
• The compiler is your friend. Each compile-time
error is potentially one less runtime or logical
error you'd have to chase down.
• Ideally, wrong code should not compile. (Static
assertion, type traits, template
metaprogramming, type system design, etc.)

Build System (2/4)
• Code build time is important. Make it as
short as possible.
– Plan the dependencies (inside your code.)
– Use precompiled headers.
– Modularize your code (into libraries.)
– Use abstract interfaces (but consider the
performance costs of polymorphism.)
– Eliminate unnecessary inter-class
dependencies.
– Use forward definitions.

Build System (3/4)
• Build variations
– Full debug
– Optimized debug
– Profiled release
– Developer release (with debug information)
– Retail
– and a couple more...

Build System (4/4)
• Strive to always have a build that is
always:
– Playable
– Performant
– Progressing
– In Memory
– Stable

Decisions to be Made
• Coding Style
• To OOP or not to OOP...
• Exceptions
• RTTI
• Templates vs. Polymorphism
• Smart pointers vs. Handles vs. Raw pointers
• RAII
• Lifetime and Ownership Management (and a bag full of
problems)
• Intrusive code instrumentation
• String format
• and much much much more...

Love Thy Compiler (1/2)
• Use EVERY little obscure compiler and platform facility
that can possibly help.
• But wrap them up, even if you don't care about portability
(in case of a change in compilers/libraries, or if you want
to disable a feature, etc.)
• Don't be afraid of assembly. Learn to read through what
your compiler produces. Love doesn't necessarily imply
trust!
• Memory is NOT fast (except on an SPU.) LOVE YOUR
CACHE TOO. Even more than your compiler.

Love Thy Compiler (2/2)
• Learn all about compiler #pragmas and intrinsics.
• Use intrinsics instead of inline assembly, because asm
usually hinders compiler optimizations.
• Learn about pointer aliasing, restricted pointers and
every other hint that you can give the compiler to
produce better (faster) code.
• Don't be afraid of assembly. Sometimes compilers
generate really bad code. Hand-crafted assembly is and
always will be an option.
• ...but not for the faint of heart. Don't drop to assembly
lightly, specially on PC.

Type System (1/3)
• Consider implementing your own containers (STL is
too frivolous with memory.)
• Type metadata and reflection
• Try not to use common type keywords (int, short,
float, etc.) Typedef them all.
– i32, u32, i16, u64, i8, u8, byte, ch8, ch16, ...
– f32, f64, v128
• ..or use stdint.h, which is a standard header but
missing from Visual C++ (still!) There are free and
conforming implementations available though.

Type System (2/3)
• Types that are equal physically, but not
semantically:
– Use different classes with no assignment
operators (or at least correct ones.) For example:
Degree vs. Radian, NormalizedVector3 vs.
Vector3
– Hungarian Notation (commonly misapplied, e.g.
in Win32 headers and docs!)
• degAngle = radAngle; is most probably
wrong and the good thing is that it indeed
looks WRONG.

Type System (3/3)
• Strings
– ASCII is dead (mostly.)
– At least differentiate between these usages:
• File names and paths.
• User input and text shown to the user.
• Internal entity names, IDs and the like.
• Text file and script contents, etc.
• And many more.
– Realize that some strings are (mostly) constant during
their lifetime. (Optimization opportunity!)

Pointers
• Don't use T* to represent a pointer to T. At least
typedef something like TPtr and use that instead (or
Ptr<T>.)
• Don't use T* to represent an array. Use a struct with
size information (templated for arrays with known sizes
at compile-time, or a member size_t (or smaller data if
suitable) if not.)
• NULL is not the only invalid pointer value (0xcdcdcdcd,
0xcafebabe, 0xdeadc0de, …) Don't use comparison
with NULL to detect this. Use a macro.
• Don't just use NULL for passing optional values. You can
use boost::optional, or at the very least abstract out
the concept.

Standard Library Independence
• File Handling and asynchronous I/O
• Directory handling
• Random numbers
• Math
• Time handling
• ...

Debugging
Consider the idea that the purpose of
writing code is not running it. Write code
for debugging.
• Logging
• Assertion
• Context metadata
• Crash handling and crash dump
• Visualizing data

Debug: Logging
• Do NOT overwhelm!
• Collect as much as possible, show as little as
needed.
• A good logging library should be:
– runtime controllable
– able to send logs to multiple targets
– able to filter what is sent where based on audience,
source, target, etc.
– fast, fast, fast
– as non-intrusive as possible
– Employed from day 0.

Debug: Assertion
• You do use them, don't you?!
• Never use the standard assert!
• Assert whenever you can, but nowhere
you shouldn't. (Hint: don't assert when a
configuration file is not found; make a new
one!)

Debug: Context metadata
• Should be totally and partially disable-able.
• Including, but not limited to:
– Machine name, user name, code revision, build number, build
parameters, time, environment, frame number, etc.
– Where does the execution flow come from? (Stack trace)
– Where are we now? File, line, function, instance.
– Which subsystem called this piece of code? Graphics? AI?
Animation? Sound?
– Why did it call this? Was it initializing? Loading data? Looking for
a collision? Walking the scene graph? Visiting AI actors?
– How long did the function take to execute?
– How Many times was it called in this frame?
– How many memory allocations did it perform?
– … (use your imagination!)

Debug: Crash handling and crash
dump
• If you have to fail, fail as soon as possible and
as loud as possible.
• Gather as much data as you possibly can: place,
time, what happened before, what was going to
happen after, context, hardware environment,
software environment, user settings, etc.
• Use a “Black Box” type construct. Collect some
data all the time, but store it only when crashing.
• Consider automatically submitting this info to
your issue tracker (during development) or
emailing it to your team (after release.)

Debug: Visualizing Data
• There is much more going on in a game than
graphics (which can be seen already!)
– Lots of data involved.
– No clear picture from mere numbers.
– Hard to understand in a single slice of time.
• Add visual representations for non-visual
objects: physics, bounding volumes, AI
algorithms, behavioral decisions, etc.
• Real-time graphs and bars too!

Memory
• NEVER use naked new/delete in real code.
• NEVER EVER use naked new/delete in real code!
• Write your own memory manager, even if it is just a
wrapper over new/delete.
• Don't forget malloc() and friends.
• Don't forget third-party libraries.
• Always be conscious about who uses how much
memory. Always.
• Always override the memory management in third party
libraries (including STL) unless you have reason not to.
• Writing a memory bug detector is quite easy. Write one
or find one.
• Memory allocation is usually a source of performance
bottlenecks as well as bugs.

In-game console
• Useful for debugging and interactive
experimentation.
• Tied-in to the scripting system.
• Make it like a good OS shell.
• Can think about it as a higher-level debugger
(inter-object relationships, etc.)
• Give access to everything you can, but consider
that the designers and artists will have to use it
too.

Serialization
• Useful for much more than saving/loading or
network communication.
• Template magic? Some form of IDL?
Übermacros?
• Consider two interchangeable formats: one
optimized binary and one human readable.
• Make everything you can serializable.
• Consider interactions with source control.
• Plan for evolution of data structures over the
course of development.

Runtime Asset management
• Everything other than code that
constitutes the game.
• Consider going beyond simple files.
• Consider treating asset management
more like writing a DBMS, than a file
system.
• Consider integrating asset fingerprinting,
change management, versioning, change
detection, hot reloading, etc.
• Consider asynchronous operations.

Time
• Do NOT underestimate time management.
• Never just use direct calls to OS to get the time for game
subsystems.
• Familiarize yourself with higher-resolution timers
(RDTSC, QueryPerformanceCounter(),
gettimeofday(), ...)
• Familiarize yourself with multi-core issues, power
management issues, Windows issues(!), user issues,...
• Consider the difference between wall clock, game clock,
camera clock, level clock, animation clock, etc.
• The more time you put in time, the better the times you'll
have when the time of debugging time comes!

Screen capture and Game Replay
• Capturing still images is a must.
• Capturing video is a big plus.
• Capturing metadata (DEM files, anyone?) to be
able to replay a scenario is a huge plus.
• The game replay subsystem can be
implemented with varying levels of complexity
and functionality. Even a simple system can
have considerable benefits.

Runtime Statistics
• Please invest in this!
• You must know your own game. This is
one very good way.
• Anything from frame-rate and tri-count to
number of memory allocations and file
accesses per frame, per second or per-
level.
• Find interesting ways to display these
data.

Automated Testing
• Full unit testing is the Holy Grail!
• Testing of games and features is usually very
time consuming. Any automation helps.
• Any change in code or assets can have
undesired side-effects: regression testing.
• You can even have automatic screen-shot
comparisons!
• Auto test non-visual parts, if you can't manage
more sophisticated methods.

Questions?

"I know that you believe that you
understood what you think I said, but I
am not sure you realize that what you
heard is not what I meant.”

We are hiring!

jobs@fanafzar.com

Preventing Complexity in Game Programming

More Related Content

What's hot (20)

Similar to Preventing Complexity in Game Programming (20)

Preventing Complexity in Game Programming

Editor's Notes