Slide 1: The Next Mainstream Programming Language: A Game Developer’s Perspective Tim Sweeney Epic Games
Slide 2: Outline  Game Development – Typical Process  What’s in a game? – Game Simulation – Numeric Computation – Shading  Where are today’s languages failing? – Concurrency – Reliability
Slide 3: Game Development
Slide 4: Game Development: Gears of War  Resources – – – – ~10 programmers ~20 artists ~24 month development cycle ~$10M budget  Software Dependencies – 1 middleware game engine – ~20 middleware libraries – OS graphics APIs, sound, input, etc
Slide 5: Software Dependencies Gears of War Gameplay Code ~250,000 lines C++, script code Unreal Engine 3 Middleware Game Engine ~250,000 lines C++ code DirectX Graphics OpenAL Audio Ogg Vorbis Music Codec Speex Speech Codec wx Widgets Window Library ZLib Data Compression …
Slide 6: Game Development: Platforms  The typical Unreal Engine 3 game will ship on: – Xbox 360 – PlayStation 3 – Windows  Some will also ship on: – Linux – MacOS
Slide 7: What’s in a game? The obvious:  Rendering  Pixel shading  Physics simulation, collision detection  Game world simulation  Artificial intelligence, path finding But it’s not just fun and games:  Data persistence with versioning, streaming  Distributed Computing (multiplayer game simulation)  Visual content authoring tools  Scripting and compiler technology  User interfaces
Slide 8: Three Kinds of Code  Gameplay Simulation  Numeric Computation  Shading
Slide 9: Gameplay Simulation
Slide 10: Gameplay Simulation  Models the state of the game world as interacting objects evolve over time  High-level, object-oriented code  Written in C++ or scripting language  Imperative programming style  Usually garbage-collected
Slide 11: Gameplay Simulation – The Numbers  30-60 updates (frames) per second  ~1000 distinct gameplay classes – Contain imperative state – Contain member functions – Highly dynamic  ~10,000 active gameplay objects  Each time a gameplay object is updated, it typically touches 5-10 other objects
Slide 12: Numeric Computation  Algorithms: – Scene graph traversal – Physics simulation – Collision Detection – Path Finding – Sound Propagation  Low-level, high-performance code  Written in C++ with SIMD intrinsics  Essentially functional – Transforms a small input data set to a small output data set, making use of large constant data structures.
Slide 13: Shading
Slide 14: Shading  Generates pixel and vertex attributes  Written in HLSL/CG shading language  Runs on the GPU  Inherently data-parallel – Control flow is statically known – “Embarassingly Parallel” – Current GPU’s are 16-wide to 48-wide!
Slide 15: Shading in HLSL
Slide 16: Shading – The Numbers  Game runs at 30 FPS @ 1280x720p  ~5,000 visible objects  ~10M pixels rendered per frame – Per-pixel lighting and shadowing requires multiple rendering passes per object and per-light  Typical pixel shader is ~100 instructions long  Shader FPU’s are 4-wide SIMD  ~500 GFLOPS compute power
Slide 17: Three Kinds of Code Game Simulation Languages CPU Budget Lines of Code FPU Usage 10% 250,000 0.5 GFLOPS Numeric Computation 90% 250,000 5 GFLOPS Shading CG, HLSL n/a 10,000 500 GFLOPS C++, Scripting C++
Slide 18: What are the hard problems?  Performance – When updating 10,000 objects at 60 FPS, everything is performance-sensitive  Modularity – Very important with ~10-20 middleware libraries per game  Reliability – Error-prone language / type system leads to wasted effort finding trivial bugs – Significantly impacts productivity  Concurrency – Hardware supports 6-8 threads – C++ is ill-equipped for concurrency
Slide 19: Performance
Slide 20: Performance  When updating 10,000 objects at 60 FPS, everything is performance-sensitive  But: – Productivity is just as important – Will gladly sacrifice 10% of our performance for 10% higher productivity – We never use assembly language  There is not a simple set of “hotspots” to optimize! That’s all!
Slide 21: Modularity
Slide 22: Unreal’s game framework Gameplay module Base class of gameplay objects Members package UnrealEngine; class Actor { int Health; void TakeDamage(int Amount) { Health = Health – Amount; if (Health<0) Die(); } } class Player extends Actor { string PlayerName; socket NetworkConnection; }
Slide 23: Game class hierarchy Generic Game Framework Actor Player Enemy InventoryItem Weapon Game-Specific Framework Extension Actor Player Enemy Dragon Troll InventoryItem Weapon Sword Crossbow
Slide 24: Software Frameworks  The Problem: Users of a framework need to extend the functionality of the framework’s base classes!  The workarounds: – Modify the source …and modify it again with each new version – Add references to payload classes, and dynamically cast them at runtime to the appropriate types.
Slide 25: Software Frameworks  The Problem: Users of a framework want to extend the functionality of the framework’s base classes!  The workarounds: – Modify the source – Add references to payload classes, and dynamically cast them at runtime to the appropriate types. – These are all error-prone: Can the compiler help us here? …and modify it again with each new version
Slide 26: What we would like to write… Base Framework package Engine; class Actor { int Health; … } class Player extends Actor { … } class Inventory extends Actor { … } Extended Framework Package GearsOfWar extends Engine; class Actor extends Engine.Actor { // Here we can add new members // to the base class. … } class Player extends Engine.Player { // Thus virtually inherits from // GearsOfWar.Actor … } class Gun extends GearsOfWar.Inventory { … } The basic goal: To extend an entire software framework’s class hierarchy in parallel, in an open-world system.
Slide 27: Reliability Or: If the compiler doesn’t beep, my program should work
Slide 28: Dynamic Failure in Mainstream Languages Example (C#): Given a vertex array and an index array, we read and transform the indexed vertices into a new array. Vertex[] Transform (Vertex[] Vertices, int[] Indices, Matrix m) { Vertex[] Result = new Vertex[Indices.length]; for(int i=0; i<Indices.length; i++) Result[i] = Transform(m,Vertices[Indices[i]]); return Result; }; What can possibly go wrong?
Slide 29: Dynamic Failure in Mainstream Languages May contain indices outside of the range of the Vertex array May be NULL May be NULL May be NULL Vertex[] Transform (Vertex[] Vertices, int[] Indices, Matrix m) { Vertex[] Result = new Vertex[Indices.length]; for(int i=0; i<Indices.length; i++) Result[i] = Transform(m,Vertices[Indices[i]]); return Result; }; Could dereference a null pointer Array access Will the compiler realize this can’t fail? might be out of bounds Our code is littered with runtime failure cases, Yet the compiler remains silent!
Slide 30: Dynamic Failure in Mainstream Languages Solved problems:  Random memory overwrites  Memory leaks Solveable:  Accessing arrays out-of-bounds  Dereferencing null pointers  Integer overflow  Accessing uninitialized variables 50% of the bugs in Unreal can be traced to these problems!
Slide 31: What we would like to write… An index buffer containing natural numbers less than n An array of exactly known size Universally quantify over all natural numbers Transform{n:nat}(Vertices:[n]Vertex, Indices:[]nat<n, m:Matrix):[]Vertex= for each(i in Indices) Transform(m,Vertices[i]) The only possible failure mode: Haskell-style array comprehension divergence, if the call to Transform diverges.
Slide 32: How might this work?  Dependent types int nat nat<n The Integers The Natural Numbers The Natural Numbers less than n, where n may be a variable!  Dependent functions Sum(n:nat,xs:[n]int)=.. a=Sum(3,[7,8,9]) Explicit type/value dependency between function parameters  Universal quantification Sum{n:nat}(xs:[n]int)=.. a=Sum([7,8,9])
Slide 33: How might this work?  Separating the “pointer to t” concept from the “optional value of t” concept xp:^int xo:?int xpo:?^int A pointer to an integer An optional integer An optional pointer to an integer!  Comprehensions (a la Haskell), for safely traversing and generating collections Successors(xs:[]int):[]int= foreach(x in xs) x+1
Slide 34: How might this work? A guarded casting mechanism for cases where need a safe “escape”: Here, we cast i to type of natural numbers bounded by the length of as, and bind the result to n We can only access i within this context GetElement(as:[]string, i:int):string= if(n:nat<as.length=i) as[n] else “Index Out of Bounds” If the cast fails, we execute the else-branch All potential failure must be explicitly handled, but we lose no expressiveness.
Slide 35: Analysis of the Unreal code  Usage of integer variables in Unreal: – 90% of integer variables in Unreal exist to index into arrays • 80% could be dependently-typed explicitly, guaranteeing safe array access without casting. • 10% would require casts upon array access. – The other 10% are used for: • Computing summary statistics • Encoding bit flags • Various forms of low-level hackery  “For” loops in Unreal: – 40% are functional comprehensions – 50% are functional folds
Slide 36: Accessing uninitialized variables  Can we make this work? class MyClass { const int a=c+1; const int b=7; const int c=b+1; } MyClass myvalue = new C; // What is myvalue.a? This is a frequent bug. Data structures are often rearranged, changing the initialization order.  Lessons from Haskell: – Lazy evaluation enables correct out-of-order evaluation – Accessing circularly entailed values causes thunk reentry (divergence), rather than just returning the wrong value  Lesson from Id90: Lenient evaluation is sufficient to guarantee this
Slide 37: Dynamic Failure: Conclusion Reasonable type-system extensions could statically eliminate all:  Out-of-bounds array access  Null pointer dereference  Integer overflow  Accessing of uninitialized variables See Haskell for excellent implementation of: – Comprehensions – Option types via Maybe – Non-NULL references via IORef, STRef – Out-of-order initialization
Slide 38: Integer overflow The Natural Numbers data Nat = Zero | Succ Nat Factoid: C# exposes more than 10 integer-like data types, none of which are those defined by (Pythagoras, 500BC). In the future, can we get integers right?
Slide 39: Can we get integers right? Neat Trick:    In a machine word (size 2n), encode an integer ±2n-1 or a pointer to a variable-precision integer Thus “small” integers carry no storage cost Additional access cost is ~5 CPU instructions But:  A natural number bounded so as to index into an active array is guaranteed to fit within the machine word size (the array is the proof of this!) and thus requires no special encoding. Since ~80% of integers can dependently-typed to access into an array, the amortized cost is ~1 CPU instruction per integer operation.  This could be a viable tradeoff
Slide 40: Concurrency
Slide 41: The C++/Java/C# Model: “Shared State Concurrency”  The Idea: – Any thread can modify any state at any time. – All synchronization is explicit, manual. – No compile-time verification of correctness properties: • Deadlock-free • Race-free
Slide 42: The C++/Java/C# Model: “Shared State Concurrency”  This is hard!  How we cope in Unreal Engine 3: – 1 main thread responsible for doing all work we can’t hope to safely multithread – 1 heavyweight rendering thread – A pool of 4-6 helper threads • Dynamically allocate them to simple tasks. – “Program Very Carefully!”  Huge productivity burden  Scales poorly to thread counts There must be a better way!
Slide 43: Three Kinds of Code: Revisited  Gameplay Simulation – Gratuitous use of mutable state – 10,000’s of objects must be updated – Typical object update touches 5-10 other objects  Numeric Computation – Computations are purely functional – But they use state locally during computations  Shading – Already implicitly data parallel
Slide 44: Concurrency in Shading  Look at the solution of CG/HLSL: – New programming language aimed at “Embarassingly Parallel” shader programming – Its constructs map naturally to a data-parallel implementation – Static control flow (conditionals supported via masking)
Slide 45: Concurrency in Shading Conclusion: The problem of data-parallel concurrency is effectively solved(!) “Proof”: Xbox 360 games are running with 48-wide data shader programs utilizing half a Teraflop of compute power...
Slide 46: Concurrency in Numeric Computation  These are essentially pure functional algorithms, but they operate locally on mutable state  Haskell ST, STRef solution enables encapsulating local heaps and mutability within referentially-transparent code  These are the building blocks for implicitly parallel programs  Estimate ~80% of CPU effort in Unreal can be parallelized this way In the future, we will write these algorithms using referentiallytransparent constructs.
Slide 47: Numeric Computation Example: Collision Detection A typical collision detection algorithm takes a line segment and determines when and where a point moving along that line will collide with a (constant) geometric dataset. struct vec3 { float x,y,z; }; struct hit { bool DidCollide; float Time; vec3 Location; }; hit collide(vec3 start,vec3 end); Vec3 = data Vec3 float float float Hit = data Hit float Vec3 collide :: (vec3,vec3)->Maybe Hit
Slide 48: Numeric Computation Example: Collision Detection  Since collisionCheck is effects-free, it may be executed in parallel with any other effects-free computations.  Basic idea: – The programmer supplies effect annotations to the compiler. – The compiler verifies the annotations. collide(start:Vec3,end:Vec3):?Hit print(s:string)[#imperative]:void Effectful functions require explicit annotations A pure function (the default) – Many viable implementations (Haskell’s Monadic effects, effect typing, etc) In a concurrent world, imperative is the wrong default!
Slide 49: Concurrency in Gameplay Simulation This is the hardest problem…  10,00’s of objects  Each one contains mutable state  Each one updated 30 times per second  Each update touches 5-10 other objects Manual synchronization (shared state concurrency) is hopelessly intractible here. Solutions? – Rewrite as referentially-transparent functions? – Message-passing concurrency? – Continue using the sequential, single-threaded approach?
Slide 50: Concurrency in Gameplay Simulation: Software Transactional Memory See “Composable memory transactions”; Harris, Marlow, Peyton-Jones, Herlihy The idea:  Update all objects concurrently in arbitrary order, with each update wrapped in an atomic {...} block  With 10,000’s of updates, and 5-10 objects touched per update, collisions will be low  ~2-4X STM performance overhead is acceptable: if it enables our state-intensive code to scale to many threads, it’s still a win Claim: Transactions are the only plausible solution to concurrent mutable state
Slide 51: Three Kinds of Code: Revisited Game Simulation Languages CPU Budget Lines of Code FPU Usage Parallelism 10% 250,000 0.5 GFLOPS Numeric Computation 90% 250,000 5 GFLOPS Shading CG, HLSL n/a 10,000 500 GFLOPS Implicit Data Parallelism C++, Scripting C++ Software Implicit Transactional Thread Memory Parallelism
Slide 52: Parallelism and purity Physics, collision detection, scene traversal, path finding, .. Game World State Graphics shader programs Data Parallel Subset Purely functional core Software Transactional Memory
Slide 53: Musings On the Next Maintream Programming Language
Slide 54: Musings There is a wonderful correspondence between:  Features that aid reliability  Features that enable concurrency. Example:  Outlawing runtime exceptions through dependent types – Out of bounds array access – Null pointer dereference – Integer overflow Exceptions impose sequencing constraints on concurrent execution. Dependent types and concurrency must evolve simultaneously
Slide 55: Language Implications Evaluation Strategy  Lenient evaluation is the right default.  Support lazy evaluation through explicit suspend/evaluate constructs.  Eager evaluation is an optimization the compiler may perform when it is safe to do so.
Slide 56: Language Implications Effects Model  Purely Functional is the right default  Imperative constructs are vital features that must be exposed through explicit effects-typing constructs  Exceptions are an effect Why not go one step further and define partiality as an effect, thus creating a foundational language subset suitable for proofs?
Slide 57: Performance – Language Implications Memory model – Garbage collection should be the only option Exception Model – The Java/C# “exceptions everywhere” model should be wholly abandoned • All dereference and array accesses must be statically verifyable, rather than causing sequenced exceptions – No language construct except “throw” should generate an exception
Slide 58: Syntax Requirement:  Must not scare away mainstream programmers.  Lots of options. int f{nat n}(int[] as,natrange<n> i) { return as[i]; } C Family: Least scary, but it’s a messy legacy f :: forall n::nat. ([int],nat<n) -> int f (xs,i) = xs !! i Haskell family: Quite scary :-) f{n:nat}(as:[]int,i:nat<n)=as[i] Pascal/ML family: Seems promising
Slide 59: Conclusion
Slide 60: A Brief History of Game Technology 1972 Pong (hardware) 1980 Zork (high level interpretted language) 1993 DOOM (C) 1998 Unreal (C++, Java-style scripting) 2005-6 Xbox 360, PlayStation 3 with 6-8 hardware threads 2009 Next console generation. Unification of the CPU, GPU. Massive multi-core, data parallelism, etc.
Slide 61: The Coming Crisis in Computing  By 2009, game developers will face…  CPU’s with: – 20+ cores – 80+ hardware threads – >1 TFLOP of computing power  GPU’s with general computing capabilities.  Game developers will be at the forefront.  If we are to program these devices productively, you are our only hope!
Slide 62: Questions?
Slide 63: Backup Slides
Slide 64: The Genius of Haskell  Algebraic Datatypes – Unions done right Compare to: C unions, Java union-like class hierarchies – Maybe t C/Java option types are coupled to pointer/reference types  IO, ST – With STRef, you can write a pure function that uses heaps and mutable state locally, verifyably guaranteeing that those effects remain local.
Slide 65: The Genius of Haskell  Comprehensions Sorting in C int partition(int y[], int f, int l); void quicksort(int x[], int first, int last) { int pivIndex = 0; if(first < last) { pivIndex = partition(x,first, last); quicksort(x,first,(pivIndex-1)); quicksort(x,(pivIndex+1),last); } } int partition(int y[], int f, int l) { int up,down,temp; int cc; int piv = y[f]; up = f; down = l; do { while (y[up] <= piv && up < l) { up++; } while (y[down] > piv ) { down--; } if (up < down ) { temp = y[up]; y[up] = y[down]; y[down] = temp; } } while (down > up); temp = piv; y[f] = y[down]; y[down] = piv; return down; } Sorting in Haskell sort [] = [] sort (x:xs) = sort [y | y<-xs, y<x ] ++ [x ] ++ sort [y | y<-xs, y>=x]
Slide 66: Why Haskell is Not My Favorite Programming Language  The syntax is … scary  Lazy evaluation is a costly default – But eager evaluation is too limiting – Lenient evaluation would be an interesting default  Lists are the syntactically preferred sequence type – In the absence of lazy evaluation, arrays seem preferable
Slide 67: Why Haskell is Not My Favorite Programming Language  Type inference doesn’t scale – To large hierarchies of open-world modules – To type system extensions – To system-wide error propagation f(x,y) = x+y a=f(3,”4”) … … ERROR - Cannot infer instance *** Instance : Num [Char] *** Expression : f (3,"4") Parameter mismatch paremter 2 of call to f: Expected: int Got: “4” ??? f(int x,int y) = x+y a=f(3,”4”)