Skip to content

Latest commit

 

History

History
530 lines (422 loc) · 18.6 KB

switching_from_imperative_to_functional_programming_with_games_in_Elm.md

File metadata and controls

530 lines (422 loc) · 18.6 KB

Switching from imperative to functional programming with games in Elm

Remark: The code and some of the concepts in this article are incompatible with Elm versions > 0.16. Thus the links to the code point to the old version, that is compatible with Elm 0.16. You can find the updated version of the game (for Elm 0.17) here

Imperative programming was my thing since I was a school boy. I wrote some small games and demoscene effects, and now develop software (primarily computer vision stuff) for a living. Recently it was mainly C++ and some Python I worked with. During the last year, David, a good friend and Haskell expert, tried to evangelize me with the benefits of functional programming (FP). So I read SICP and LYAH. I actually understood some parts, although not all, and solved few and very small toy problems in Scheme and Haskell. But I did not see how one could use such a programming style voluntarily, because the imperative solutions came about an order of magnitude quicker to my mind than the functional ones did.

Of course, in C++ I now used a bit more from the algorithm header and also tried some functional things in Python, yet I just did not feel how digging deeper into this paradigm could really be worth the trouble. E.g. why should I foldr over a list if I can for-loop over an array? In retrospect it perhaps also was the uncomfortable situation of again being a complete newbie that kept me from continuing.

Whatever it was, I more or less accidentally stumbled upon Elm. The examples did not look so scary like many type theory filled Haskell tutorials, so I gave it a shot and read through all the tutorials/articles while constantly playing around with the stuff I just found out.

After solving the challenges in the [pong tutorial] (http://elm-lang.org/blog/Pong.elm) I decided to write this Breakout clone, which turned out to be a very interesting and fun undertaking. Try it out! :-)

Even though some of the following things will sound naive to experienced FP developers (or even to me in some months or years), the rest of this little essay will describe my learning experience with that project, how it is structured and why it finally convinced me of the advantages of functional (reactive) programming and motivated me to immediately continue with [a second game project] (https://github.com/Dobiasd/Maze). It is not meant to serve as a full FP or game development tutorial, but perhaps it can inspire you to have a deeper look into the FP paradigm and then maybe share my excitement. :-)

Differences to a hypothetical imperative implementation in C++

If I had written this game (resp. a non browser version of it) in C++, I probably would have used SFML, which is a very good library for making games like this. I already used it to write a Snake like game. My cost estimations for that project would probably be more man-days than it surprisingly took me to do it in Elm, a language I had no experience with at all!

One reason for that is the much shorter edit/compile cycle in Elm, which reduces to just one click in your browser. But OK, except the hot-swapping, that is also possible with some imperative languages. The much more astonishing fact for me was, that I did not need many of these cycles. Sometimes I wrote code for nearly an hour and as soon as Elm's Haskell like type system did not give me errors any more while compiling, the code just worked! There was very rarely a need to debug it at all! I guess this comes from the notion, that if you look at a pure functions you just have to think about what it stands for and not what it will do to something else under certain circumstances etc. Also when just thinking in expressions and no more in statements, there is not so much control flow you have to emulate in your head. And it is easier to structure your code. The temptation to write spaghetti code functions is not that big and if one still grows too long, it is very easy to factor out the parts that can stand meaningful for their own. Furthermore the refactoring is not scary at all. I didn't introduce one single bug while factoring out stuff to add new functionality, like the traction between the paddle and the ball.

Also everything is much more concise. Just compare the two following snippets:

std::list<int> l;
for ( int i = 1; i <= 10; ++ i )
    l.push_back( i * i );
l = map ((^)2) [1..10]

Sure, what is more readable/pretty is also a matter or habit/taste. But beside the terseness there come other benefits with the abstract separation of the control structure, e.g. if you want to decide from the outside what to do with the values:

buildPairs l f = map2 (,) l <| map f l
l1 = buildPairs [1..5] ((^)2)
l2 = buildPairs [1..5] sqrt
l3 = buildPairs [1..5] ((*)2)
l4 = buildPairs [1..5] ((+)1)

You can also decide which direction you prefer to read:

la = map ((^)2) [1..10]    -- normal function application
lb = ((^)2) `map` [1..10]  -- infix notation
lc = [1..10] |> map ((^)2) -- forward application

And many design patterns involving inheritance and boilerplate code in C++ just disintegrate into thin air when you have functions as first class citizen in FP.

Don't get me wrong. I don't want to bash C++, and I still think it is a great language if it comes to performance critical system programming, and I will continue to use it for appropriate tasks. But the high control over nearly every byte in your system comes at the cost of increased development time. So every tool has its usage. You probably can get a screw into a wall using a hammer, but if you feel that a big part of your effort is wasted energy, there is presumably a better solution. ;-)

Sure, with Python things would likely have been much different compared to C++, but also there the [advantages of pure FP] (http://www.haskell.org/haskellwiki/Functional_programming#Benefits_of_functional_programming) can not always be utilized fully, and still targetting the browser so easily in a very clean and abstract way is as far is I know a unique characteristic of Elm.

Concept

OK, now we eventually come to the actual game. ;-) The pureness of Elm forces us to split it into three parts.

  • The model includes all the state of our game like the positions and speed of the objects. We also have to think about the user input we are interested in.
  • The updates part describes how the game states transition from one point in time to a subsequent one, given a certain set of inputs.
  • The view finally brings the game state onto the screen.

Let's examine how this can look in a very much simplified version of our desired game:

-- skeleton

import List exposing (map, map2)

import Graphics.Element exposing (show, Element)
import Keyboard
import Text
import Time exposing (Time, fps)
import Signal exposing (Signal, foldp)
import Signal
import Window


-- model

direction : Signal Int
direction = Signal.map .x Keyboard.arrows

type alias Input = { dir:Int, delta:Time }

input : Signal Input
input = Signal.map2 Input direction (fps 60)

type alias Positioned a = { a | x:Float }

type alias Player = Positioned {}

player : Float -> Player
player x = { x=x }

type alias Game = { player:Player }

defaultGame : Game
defaultGame = { player = player 0 }


-- updates

stepGame : Input -> Game -> Game
stepGame {dir,delta} ({player} as game) =
  let
    player' = { player | x = player.x + delta * toFloat dir }
  in
    { game | player = player' }

gameState : Signal Game
gameState = foldp stepGame defaultGame input

-- view

main : Signal Element
main = Signal.map show gameState

Model

OK, the import stuff should be clear. Now, direction is a signal that has a value of -1, 0 or 1 and updates if the user presses or releases a key. If it is not clear to you what a signal is, I suggest you first read Evan's article What is Functional Reactive Programming? before continuing here. I will wait here for you to return. =)

type alias Input = { space : Bool, dir : Int, delta : Time.Time } just tells us, that all the inputs we are interested in are the direction the user is going to with the arrow keys and a time delta. This delta holds the time passed since the last update. We aim for 60 frames per second.

For now our Player is a positioned nothing. We use the syntax for extensible records here and define a constructor for it.

Our Game just holds the players information, nothing else, and the default game has a player positioned at 0/0.

Updates

stepGame is the one function describing all the changes our game state can ever experience. Given an input and the current game state, it returns the next one. All we do for now is to update the players x position with the direction the user chooses. The direction is multiplied by delta, because we want the game to always run at the same absolute speed, even at machines that can not provide the full 60 fps.

View

At the moment the view just displays the players coordinates as text. Not very fancy, but enough to see that stepGame correctly updates our model.

Making the game a game

Changing the x value of our player with the keyboard is of course already extremely awesome, and we could spend countless hours exploring all corners of this deep new gaming concept, but at some point we want more. We want bricks and a ball to smash them into pieces.

So how do we now get from our skeleton to the final game?

First let us complete our model and our view, and then write the update code to will everything with life.

Model

In the final version our game record contains not just the player, but also the bricks still left in the game, the ball and the number of spare balls.

type alias Game =
    { state : State
    , gameBall : Ball
    , player : Player
    , bricks : List Brick
    , spareBalls : Int
    , contacts : Int
    }

It should be obvious for what the single record entries stand, and their particular types are in the source. (contacts is just used to count the number of paddle ball collisions for the overall score.)

The state is a more interesting (especially later in the update section). Our game can wait to begin (Serve), be in the actual playing phase (Play) or be over and thus be Won or Lost.

This leads to the following ADT:

type State = Play | Serve | Won | Lost

The rest is quite trivial, I guess. Please tell me if I am wrong with this assumption. ;-)

View

Most parts of the view are somewhat boring, because Elm makes all this very easy. The function display just defines how a given game state will look like on the screen. Most of the techniques used there are covered nicely in Introduction to Graphics.

More interesting is how we get the game to always use the browser windows maximally while preserving the correct aspect ratio. In the view configuration we defined (gameWidth,gameHeight) = (600,400) but on some devices this may be too small or even too big. Elm makes the scaling a charm. We define everything in out default size, but use displayFullScreen to fit it to our screen.

displayFullScreen : ( Int, Int ) -> Form -> Element
displayFullScreen ( w, h ) content =
    let
        gameScale = min (toFloat w / gameWidth) (toFloat h / gameHeight)
    in
        collage w h [ content |> scale gameScale ]

displayFullScreen just calls scale with our filled Form and the needed scale factor. This does not mean, that the game is rendered to 600x400 and then the resulting image is scaled, no. The whole game is scaled before it is actually rendered onto the screen. This means that there will be no image scaling artifacts, even on displays with scale factors far away from '1.0'. :)

Updates

The Updates are the coolest part since it is here where all the action actually happens.

Let's see, we foldp over our defaultGame with stepGame, so let's look at this.

stepGame : Input -> Game -> Game
stepGame ({ dir, delta } as input) ({ state, player } as game) =
    let
        func =
            if state == Play then
                stepPlay
            else if state == Serve then
                stepServe
            else
                stepGameOver
    in
        func input { game | player = stepPlayer delta dir player }

Since the paddle can be moved regardless of the game state, the players position is already updated here. All the other actions are state specific, so the remaining tasks are dispatched by the current state.

stepServe and stepGameOver do nothing special, so let's look at stepPlay:

stepPlay : Input -> Game -> Game
stepPlay { delta } ({ gameBall, player, bricks, spareBalls, contacts } as game) =
    let
        ballLost = gameBall.y - gameBall.r < -halfHeight

        gameOver = ballLost && spareBalls == 0

        spareBalls' =
            if ballLost then
                spareBalls - 1
            else
                spareBalls

        state' =
            if gameOver then
                Lost
            else if ballLost then
                Serve
            else if List.isEmpty bricks then
                Won
            else
                Play

        ( ( ball', bricks' ), contacts' ) =
            stepBall delta gameBall player bricks contacts
    in
        { game
            | state = state'
            , gameBall = ball'
            , bricks = bricks'
            , spareBalls =
                max 0 spareBalls'
                -- No -1 when game is lost.
            , contacts = contacts'
        }

If our ball is lost and we do not have any spare balls left, the game is over. Simple. stepBall seems to be the place where the collision is handled. It's type already tells us a lot about it: stepBall : Time -> Ball -> Player -> [Brick] -> Int -> ((Ball,[Brick]), Int) Using a time delta, it takes values of the ball, player and bricks and returns new values for them. The number of paddle ball contacts may also be increased.

stepBall : Time.Time -> Ball -> Player -> List Brick -> Int -> ( ( Ball, List Brick ), Int )
stepBall t ({ x, y, vx, vy } as ball) p bricks contacts =
    let
        hitPlayer = (ball `within` p)

        contacts' =
            if hitPlayer then
                contacts + 1
            else
                contacts

        newVx =
            if hitPlayer then
                weightedAvg [ p.vx, vx ] [ traction, 1 - traction ]
            else
                stepV vx (x < (ball.r - halfWidth)) (x > halfWidth - ball.r)

        hitCeiling = (y > halfHeight - ball.r)

        ball' =
            stepObj
                t
                { ball
                    | vx = newVx
                    , vy = stepV vy hitPlayer hitCeiling
                }
    in
        ( List.foldr goBrickHits ( ball', [] ) bricks, contacts' )

First it checks for paddle player collisions and updates the ball's velocity and the contact count accordingly.

The return type (((Ball,List Brick), Int)) may look somewhat odd at first glance, but thanks to pattern matching it did not pose a problem for the caller (stepPlay): ((ball', bricks'), contacts') = stepBall ... And it renders itself quite handy when looking at (foldr goBrickHits (ball',[]) bricks, contacts')

OK, what does foldr goBrickHits (ball',[]) do? I will not explain the behaviour of [folds in general] (http://www.haskell.org/haskellwiki/Fold#List_folds_as_structural_transformations) here, but in our case the used function goBrickHits takes one single brick at a time and the accumulator, which we initially set to (ball',[]). It contains the already changed ball and at first no bricks at all:

goBrickHits : Brick -> ( Ball, List Brick ) -> ( Ball, List Brick )
goBrickHits brick ( ball, bricks ) =
    let
        hit = ball `within` brick

        bricks' =
            if hit then
                bricks
            else
                brick :: bricks

        ball' =
            if hit then
                speedUp { ball | vy = -ball.vy }
            else
                ball
    in
        ( ball', bricks' )

Initially it checks if the ball is colliding with the current brick, and if this is true we do not put this brick in our accumulator since it just was destroyed by the ball. We also invert the balls vertical speed. If no hit occurred, we just cons the brick to the bricks that are still in the game. Et voilà. Who needs for-loops anymore? :-)


OK, that's it. Since at the moment of writing I'm quite new to all this, I guess there is still much room for improvement of the code and this article. If you have suggestions please let me know. :-)

Conclusion

Writing Breakout in Elm was surprisingly easy and a lot of fun. It was the one experience I needed to get practice and thus gain confidence in programming purely functional. I am really looking forward to learning more about Elm (and FP in general) and how this awesome language will develop during the next years.

And David, you were right. FP really rocks. ;-)

Discuss on reddit or HN