luminance-0.1.1: Type-safe, dependently-typed and stateless graphics framework

Copyright(C) 2015 Dimitri Sabadie
MaintainerDimitri Sabadie <>
Safe HaskellNone



What is luminance?

luminance is a small yet powerful graphics API. It was designed so that people can quickly get their feet wet and start having fun with graphics in Haskell. The main idea is to unleash the graphical and visual properties of GPUs in a stateless and type-safe way.

This library doesn’t expose any architectural patterns or designs. It’s up to you to design your program as you want and following your own plans. Because it’s a graphics and rendering API, you won’t find several common things you find in animations, games or simulations. If you need those you’ll have to look for dedicated libraries instead.

One of the most important thing you have to keep in mind is the fact that luminance won’t provide you with anything else than working with the GPU. That is, it won’t even provide functions to open windows. That’s actually a good thing, because then you’ll be able to use it with any kind of windowing and system library you want to!

The drawback is about safety. If you screw up setting up the OpenGL context, there’s no way luminance will work. If users feel the need, a few dedicated packages will be uploaded, like luminance-glfw to add GLFW-b support for instance.

Getting started

Setting up the window and OpenGL context

The first thing to do is to create a window. Here’s a typical GLFW-b snippet to create such a window.

  initialized <- init
  when initialized $ do

In the first place, we initialize the GLFW-b library and make sure everything ran smoothly.

    windowHint (WindowHint'Resizable False)
    windowHint (WindowHint'ContextVersionMajor 4)
    windowHint (WindowHint'ContextVersionMinor 5)
    windowHint (WindowHint'OpenGLForwardCompat False)
    windowHint (WindowHint'OpenGLProfile OpenGLProfile'Core)
    window <- createWindow 800 600 "luminance application" Nothing Nothing

That part just setup the window’s OpenGL hints so that we create a compatible context for luminance. luminance will work with OpenGL 4.5 only, don’t even try to make it work with a lower implementation. We also disable the forward compatibility because we don’t need it and ask to stick to a core profile.

    case window of
      Just window' -> do
        makeContextCurrent window
        swapInterval 1
        -- we’re good to go!
        destroyWindow window'
      Nothing -> hPutStrLn stderr "unable to create window; please check your hardware support OpenGL4.5"

We then test window. If we have successfully opened the window, we go on by making the OpenGL context of the window the current one for the current thread, set the swap interval (that’s not important for the purpose of that tutorial) and we’re good to go. Otherwise, we just display an error message and quit.


We finally close the GLFW-b context to cleanup everything.

Preparing the environment for luminance

A lot of the functions you’ll use work in special types. For instance, a lot of create* functions will require (MonadIO m,MonadResource m,MonadError e m) => m or so. For that reason, we’ll be using a type of our own and will unwrap it so that we end up in IO in the end.

Here’s the type:

  type App = ExceptT AppError (ResourceT IO)

  newtype AppError = AppError String deriving (Eq,Show)

And we’ll unwrap from our type to IO with:

  runResourceT . runExcepT

Getting something to the screen

About the screen

luminance generalizes OpenGL concepts so that they’re made safer. In order to render something onto the screen, you have to understand what the screen truly is. It’s actually… a back buffer – assuming we have double buffering enabled, which the case with GLFW-b by default. So rendering to the screen is the same thing than rendering to the back buffer and ask GLFW-b to swap the back buffer with the front buffer.

And guess what. luminance wraps the back buffer into a Framebuffer object. You can access it through defaultFrambuffer. That value will always represent the back buffer.

About batched rendering

In most graphics frameworks, rendering is the act of taking an object and getting it rendered. luminance follows a different path. Because of real world needs and, well, real applications, you cannot do that in luminance. Because, what serious application will render only one object? None. If so, then it’s an exception. We shouldn’t design our libraries and interface for the exceptions. We should build them for the most used case, which is, having a lot of objects in a scene.

That’s why luminance exposes the concept of batched rendering. The idea is that you have to gather you objects in batches and render them all at once. That enables a correct sharing of resources – for instance, framebuffers or textures – and is very straight-forward to reason about.

luminance has several types of batches, each for the type of shared information. You can – up to now – shared two information between the rendered objects:

  • framebuffer: that means you can create a FBBatch that will gather several values under the same Framebuffer;
  • or shaders: that means you can create a SPBatch that will gather several values under the same shader Program.

The idea is that the SPBatches are stored in FBBatches. That creates a structure similar to an AST luminance knows how to dispatch to the GPU.

About shader stages

luminance supports five kinds of shader stage:

  • tessellation evaluation shader
  • tessellation control shader
  • vertex shader
  • geometry shader
  • fragment shader

Additionnaly, you can create compute shaders but they’re not usable up to now.

When creating a new shader, you have to pass a String representing the source code. This will change in the end. An EDSL is planned to make things easier and safer, but in the waiting, you are stuck with String, I’m sorry.

You have to write GLSL450-conformant code.

About uniforms

Shaders are customized through uniforms. Those are very handy and very simple to use in luminance. You have the possibility to get them when creating shader Programs. The createProgram function expects two parameters: a list of shader Stages and a uniform interface builder function. That function takes another function as parameter you can use to retrieve a uniform U by passing Either a String for the name of the uniform or a Natural for its explicit semantic. Be careful when using explicit semantics though; they’re not tested.

Here’s an exemple of such a use:

  (program,uniformInterface) createProgram shaderStages $ uni - do
    resolutionU <- uni $ Left "resolution"
    timeU <- uni $ Left "time"
    pure $ divided resolutionU timeU

In that example, uniformInterface has type U ((Float,Float),Float), (Float,Float being the type of the resolutionU part and Float being the part for timeU. divided is a method of Divisible – the typeclass of divisible contravariant functors – which is defined in the "contravariant" package.

If you don’t need uniform interface, you can build a dummy object, like (), or simply use the appropriate createProgram_ function.

About RenderCmd and Geometry

RenderCmd is a very simple type yet powerful one. It’s a way to add stateless support to OpenGL render commands – draw commands, actually. It gathers several information you can set when performing a draw command. A RenderCmd can hold any type of object, but the most useful version of it holds Geometry.

A Geometry is a GPU version of a mesh. It’s composed of vertices, indices and a primitive mode used to know how to link vertices between each others. Sometimes, Geometry doesn’t have indices. That’s called direct geometry, because the vertices are supposed to be directly used when creating primitives. If you use indices, then you have a indiced geometry and the vertices can linked by looking at the indices you’ve fed in.

A Geometry is created with the createGeometry function and a RenderCmd is created with renderCmd. You’re supposed to create a Geometry once – while loading your resources for example – and the RenderCmd can be created on the fly – it doesn’t require IO.

Putting all together

Let’s draw a triangle on the screen! First, we need the vertices!

  vertices :: [V 2 Float]
  vertices =
      V2 (-0.5) (-0.5)
    , V2 0 0.5
    , V2 0.5 (-0.5)

V 2 is a cool type used to represent vertex attributes. You’ll need DataKinds to be able to use it.

Then, we don’t need indices because we can directly issue a draw. Let’s then have the GPU version of those vertices:

  triangle <- createGeometry vertices Nothing Triangle

Then, we need a shader! Let’s write the vertex shader first:

  in vec2 co;
  out vec4 vertexColor;
  vec4 color[3] = vec4[](
      vec4(1., 0., 0., 1.)
    , vec4(0., 1., 0., 1.)
    , vec4(0., 0., 1., 1.)
  void main() {
    gl_Position = vec4(co, 0., 1.);
    vertexColor = color[gl_VertexID];

Nothing fancy, except that we pass vertexColor to the next stage so that we can blend between vertices.

Now, a fragment shader:

  in vec4 vertexColor;
  out vec4 frag;
  void main() {
    frag = vertexColor;

Now, let’s create the shader Stages and the shader Program:

  program 'sequenceA' ['createVertexShader' vsSrc,'createFragmentShader' fsSrc]>= createProgram_

Once again, that’s pretty straight-forward.

Finally, we need the batches. We’ll need one FBBatch and one SPBatch.

  let spb = shaderProgramBatch_ program [stdRenderCmd_ triangle]
      fbb = framebufferBatch defaultFramebuffer [anySPBatch spb]

Ok, so let’s explain all of this. shaderProgramBatch_ is a shorter version of shaderProgramBatch you can use to build SPBatch. The extra underscore means you don’t want no uniform interface. We pass our program and a singleton list containing a RenderCmd we create with the stdRenderCmd_. Once again, the extra underscore stands for no uniform interface. We then just pass our triangle. Notice that both stdRenderCmd and stdRenderCmd_ disable color blending and enable depth test so that you don’t have to pass those information around.

Then, we create the FBBatch. That is done via the framebufferBatch function. It takes the Framebuffer to render into – in our case, the defaultFramebuffer, which is the back buffer. We also pass a singleton list of the universally quantified SPBatch with the anySPBatch function.

We just need to issue a command to the GPU to render our triangle. That is done with a constrained type, Cmd.

  void . runCmd $ draw fbb

We don’t need the result of runCmd in our case so we discard it with void. runCmd runs in MonadIO.

We just need to swap the buffers with swapBuffers window – see GLFW-b for further details – and we’re good!

Dealing with Texture2D

Up to now, luminance only supports 2D-textures. More texture types will be added as luminance gets mature. The interface might change a lot, because it might be very inefficient, especially when converting from containers to others.