Table of Contents

tl;dr This is part of a series where we’ll be writing a roguelike using FRP and Haskell. This first article is about setting up the main loop.


I’d like to learn how to write a push based FRP implementation and so I decided I would first take a stab at flipping my pull based FRP into a pushy one. My FRP is called varying. It’s very simple and inspired by netwire. It uses automatons to generate a stream of varying values - hence the name. For info on the core concepts of varying check out the hackage docs. varying is a rather squishy FRP (and I use the term FRP liberally) but it’s fun and simple. In this article we’ll be building a quick demo ‘game’ to demonstrate how to set up the infrastructure needed for a bigger application.

Get the Code

This is a Literate Haskell file which can be downloaded from the github repo. To build, run stack build from the project directory. Go here for help with stack.


Our first import is varying, which allows us to describe values that change over time and user input.

Next up is the graphics library gelatin-picture, which we use to describe two dimensional pictures. Since we’ll be rendering for desktop with glfw we will use gelatin-glfw. It’s currently the only backend for gelatin.
gelatin-glfw re-exports GLFW-b and gelatin-picture so we don’t have to clutter our workspace with those imports.

Next we’ll need some infrastructure in the form of TVar and WriterT. We’ll use TVars to synchronize updates across threads and WriterT to allow our network entities to reach out to the world and each other.

Lastly we’ll need some miscellaneous bits and pieces.


We need to be able to describe our game and since this is Haskell we’ll use lots of types.

UserInput will represent everything we want to push into our FRP network. If our game or display logic needs to know about it, it should be covered by UserInput.

We use OutputEvent with WriterT to allow entities within our network to have a very managed effect on the network as a whole.

We’ll be rendering Pictures from gelatin-picture. We’ll talk more about rendering in the [rendering][#rendering] section. Here we just do a litty type synonym to keep us from having to type “()” all over the place.

The Network is a varying value. This means that it represents a value that changes over some domain. When you see the type of a varying value as VarT m a b it means that an output value b varies over input a within an effect m. varying is an arrowized FRP implementation with a twist, so if you’ve ever used netwire you’ll be a bit familiar. Some differences between varying and netwire are

  • varying’s inhibition is explicit using the type VarT m a (Event b).
  • varying’s time is not encoded in its type.

Here our Network is defined as a varying value that depends on UserInputs and produces a Pic inside the Effect monad. The Effect monad allows our network streams to call out to the rest of the world, writing attempted side-effects to the network as a whole through our WriterT monad stack.

A Bit More Infrastructure

Finally we can declare our main infrastructure type, AppData. An AppData contains the Network in its current state, a cache of renderers that will be managed by renderable, and a list of user input events.

The Network

We need a network to test our infrastructure, so for now we’ll use the simplest network I can think of that demonstrates change over time and user input - a circle that follows the mouse, changing shape and color over time.

Cursor Move Events

In order to do this we’ll first need a stream of cursor move events. We’ll be using V2 Float from linear to represent our points.

This type signature shows that our stream will ‘consume’ user input and ‘emit’ position events.

For the implementation we use the var constructor to turn a pure function into a stream

and then combine that with the event generator onJust which takes a Maybe a as input and produces a stream of Event a. We also use the plug right ~> combinator to plug the output of var f into the input of onJust.

And now we write our function that maps input values to Maybe (V2 Float).

Cursor Position

Next we need a stream that produces the current cursor position each frame. Until the cursor moves we need a default cursor position. V2 -1 -1 seems pretty good.

foldStream works just like foldl, but it only operates on streams of Events, folding event values into an accumulator. In this case the accumulator is simply the latest cursor position.

Time Deltas

To demonstrate that our pull network can be run on demand we’ll describe a circle that follows the cursor and changes shape and color over time. Eventually the circle will stop changing shape and color so unless there’s an input event we should see the network go to sleep.

So far we have a stream for the cursor position but now we’ll need a stream of time deltas. The reason we need deltas is that we’ll use some tweening streams from varying that require deltas as input. This breaks some of the rules of hard FRP - we’re not supposed to deal in deltas, but I think you’ll find that once it’s done we don’t have to use deltas for anything more than a plug.

We need to fill in the gaps in timeUpdated when it doesn’t produce an event.

The implementation of deltas says "produce 0 unless timeUpdated produces an event - if so, produce the value of that event.

Requesting Updates

In order to have smooth animation over time we need to know that the network requires frequent updates. In our main loop we’ll block until user input happens or until the network requests that itself be updated.

Using the monadic constructor varM we write a stream that can reach out to the Writer monad, make the request and pass whatever input it received through as output. It’s a bit hacky as far as FRP goes (we’re only using its side-effect), but this is what’s required to use our system in a pushy fashion.

You can see that both var and varM use regular functions to create a stream. The input to the stream is the only parameter to the function and the output of the function becomes the output of the stream. Check out the other constructors in the varying docs.


Combining our deltas and requestUpdate streams gives us our main time stream. Whenever a part of our network depends on time the underlying infrastructure should receive a request that the network be updated. If no part of the network depends on time our main loop will sleep and only wake to render on user input.

Since we now have time flowing through our network we can use the tweening capabilities that ship with varying in Control.Varying.Tween. Tweens run one step higher in abstraction in order to play with varying values that are only defined over a select domain. These are called splines.

Tweening With Splines

A spline is a temporary changing value that has an end result. Splines are essentially chains of event streams. When an event stream stops producing the current spline terminates and the next event stream takes over. The input, output and result value of a spline are encoded in the spline’s type signature. A ‘Spline a b m c’ describes a spline that takes ‘a’s as input, produces ’b’s as output - runs in the ’m’ monad and results in ‘c’. Our tweening spline will take time (Float) as input, give Float as output and result in the last tweened output value, also Float. We’ll want to be able to change the duration of the spline so we’ll write a function that takes the duration and returns our spline.

Above we tween from 1 to 0 over t/2 seconds using an easing function. The spline produces the interpolated values until t/2 seconds, and then results in the last interpolated value, which is either 0 or very close to 0.

Now we complete the tween.

As you can see, with splines we can use monadic notation. Since a spline is only defined over a certain domain it can be considered to terminate, giving an end result and allowing you to chain another spline. This chaining or sequencing is what makes splines so useful for defining a signals behavior over events.

Actually Using Splines

Our tweening spline represents a number over time but splines are only continuous over a finite domain, which is in this case t seconds. We’d like to use a completely continuous signal (a Var m InputEvent Pic) to tween our Pic. Spline runs at one level of abstraction above VarT and as such, we can create an output stream of the spline, turning it back into a VarT. All we need is to use outputStream along with an initial value. The resulting stream will produce the values of the spline until the it concludes. Once the spline terminates the stream will repeat the last known value forever. If the spline never produces, the initial value will be produced forever.

We’d like to demonstrate that the network sleeps when time is no longer a dependency, which means we’ll have to set up a network that depends on time for a while and then moves on. Since this sounds like sequencing event streams we’ll use splines again.

Note how the type signature now contains Effect since we’ll need to use time, which requires the ability to access the Writer monad to write out update requests.

Above we plug time into two streams - the first is our tweening stream and the second is an event stream we use as a timer. time ~> after t will produce Event () forever after t seconds. We combine the two streams using the combinator untilEvent. This combinator produces output values using the first stream until an event occurs in the second stream, then returns the last values of each in a tuple. after will produce Event (), and untilEvent unwraps the event value for you, tupling it up - making the right value of our spline’s result (), which we can ignore.

All together, multSequence is a spline that depends on time for t seconds and then returns 1 forever. This should help us demonstrate that the network no longer depends on time and can render sporatically whenever a user event comes down the pipe.

Lastly we’ll turn that spline into a continuous stream.

The Big Picture

Now we can combine our network. We start by writing a pure function that takes a position, scale, red, green and blue parameters and returns a Pic. This Pic is what we’ll render each frame.

Here we use functions from gelatin-picture to translate, scale and fill a circle of radius 100.

We put it all together with varyings Applicative instance to construct our Pic stream.

What this bit of code says is that our network is a picture that follows the cursor position changing scale over three seconds, changing its red color channel over one second, its green over two seconds and its blue over three seconds. This will all happen in parallel so all time-based animation should conclude after three seconds. At that point the circle will be white (multOverTime ends on 1.0), with a radius of 100 and following the cursor.

Our Game Loop

In our main loop we need to make a window and write functions for pushing input into our network. We’ll also sample our network and render it.

Start up glfw, receiving a Rez. A Rez is a type of resource that gelatin-glfw uses to render. It’s a composite type containing a glfw window and some shaders.

Next we’ll need a TVar to contain our app data - that way we can access and modify it from any thread. This is going to be a big part of turning our pull based FRP network into a push-esque system, where we only render what and when we need.

varying is a pull based FRP implementation. This roughly means that as the programmer you describe a network using various streams, combining them together until you have one final stream that you can “pull” or sample from each frame. Your per-frame game data will essentially just fall out of the network every frame. The downside to a pull network is that you have to sample it often, typically every frame - regardless of whether or not anything has changed. This is what we are trying to avoid.

In an attempt to remedy that situation we’ll only run the network when we receive an event from glfw - or if our network requests that we run it. Let’s define the function we’ll use to “push” a user event into our app.

That’s it! We simply use atomically $ modifyTVar' to update the app’s appEvent slot, adding the new input at the end of the list. Haskell makes threading suuuuper easy.

Separately we need a function to sample and render. We’ll run this any time glfw receives an event of any kind. Actually we’ll run this any time glfw wakes the main thread but that’s almost the same thing.

Above we print the current time just so we know the app is stepping in the console. We can use this to verify that we’re only running the step function when an event occurs.

The Network Step

Now we’ll run the network. We keep track of the last time we ran the network and create a new time InputEvent to step our time based streams with, then we add that on to the end of our stored events and sample the network using stepMany.

stepMany takes a list of input to iterate your network over, returning the last sample. Earlier I mentioned that UserInput needs a Monoid instance. The reason for this is that stepMany runs over each item of the list until it gets to the empty list. It then runs one more time using the mempty event. Ultimately, if you fed your network an empty list of events, stepMany would create one and step your network over it. If that sounds funky don’t worry, there are other strategies to use for sampling, but this makes the most sense here.

Now we can render our Pic.

And write our new app state, making sure to clear out our events.

Then apply our network’s requests. We fold our output using a Set since we only want unique requests. We don’t want time going super fast just because more network nodes request it. Time…woah.

Here’s where the update request magic happens. We spawn a new thread to wait a duration, whatever we see fit. In this case it’s oneFrame, or one thirtieth of a second. Then we push an InputTime event into our app and call glfw’s postEmptyEvent. postEmptyEvent will wake up the main thread from waitEvents, which you’ll see later.

But what about our user input events? For this we can wire up glfw’s nifty callbacks. They’ll simply push some events in our event queue.

Next is a nifty hack. In this callback we won’t push an event, we’ll just call step to update our app while the window is refreshing. This gives the app the ability to run even when the window is actively being resized.

And now we can loop forever! We process all the stored events and render our app. Then we call waitEvents which will put the main thread to sleep until some events have been stored - rinse and repeat.

Output requests for updates will be scheduled in a separate thread and will post an empty event, waking up the main thread from waitEvents - causing our loop to recurse.


FRP is a pretty cool thing. It’s got some great ideas and its a nice way to organize your code. It encourages very granular functions and by providing a small feedback mechanism you can remedy some of the bitter taste of constant rendering.

Hopefully this tutorial has been helpful. Please comment at HN or Reddit, constructive or not! You can say things like “I hate me an FRP” or “I learn me some streams, hot dang.” Thanks for reading :)

Oh - also - don’t forget that this is a literate haskell file and can be built and run in standard cabal or stack fasion.