Skip to content

Using MetalKit part 1

The MetalKit framework was announced at the WWDC 2015 and it brings a great deal of improvements and new features for Metal. Meet MTKView, a subclass of NSView/UIView. It comes with an embedded Metal layer, and it also manages the framebuffer and its render target attachments, as well as takes care of the draw loop.

Let’s create a new Cocoa Application (since iOS simulators do not support Metal). Make sure only the Swift and Use Storyboards fields are selected. Next, let’s create a new class named MetalView.swift of type NSView (for now). Once created, go to the storyboard and select the Viewunder the View Controller as set its class to a MetalView type under Identity Inspector as seen in the image below.

Do the same for the View Controller and in Identity Inspector under Class delete View Controller because we’re not going to use it. Also delete ViewController.swift since we don’t need it anymore. Now go back to MetalView.swift and type import MetalKit. There are two ways we can prepare our class for drawing: either conform to the MTKViewDelegate protocol and implement its drawInView(:) method, or subclass MTKView and override its drawRect(:)method. We choose the latter, so go ahead and change the class type from NSView to MTKView, and create a new method named render() that has the following content:

func render() {
    let device = MTLCreateSystemDefaultDevice()!
    self.device = device
    let rpd = MTLRenderPassDescriptor()
    let bleen = MTLClearColor(red: 0, green: 0.5, blue: 0.5, alpha: 1)
    rpd.colorAttachments[0].texture = currentDrawable!.texture
    rpd.colorAttachments[0].clearColor = bleen
    rpd.colorAttachments[0].loadAction = .Clear
    let commandQueue = device.newCommandQueue()
    let commandBuffer = commandQueue.commandBuffer()
    let encoder = commandBuffer.renderCommandEncoderWithDescriptor(rpd)

Let’s walk over the code line by line. First, we create a device. We set this device as our view’s device or otherwise it will stay nil and will crash the app. Optionally, we can modify the view’s drawable properties before drawing. Next, we create a render pass descriptor which allows us to configure a render pass with the current drawable’s texture attached as its primary color. For fun, we create a nice color named bleen which is half blue and half green. Next, we create a command queue for the device, and use the queue to create a command buffer. Finally, we use the command buffer to create the render command encoder to perform the draw calls. For each iteration of the draw loop, a new MTLRenderPassDescriptor object is made available when queried from currentRenderPassDescriptor. This object is created based on the currentDrawable object. Presentation is not handled by MTKView so it is our responsibility to first check that both currentRenderPassDescriptor and currentDrawable are not nil before we call the presentDrawable(:) method on the view’s current drawable.

Let’s refer to the Metal documentation for more details. The Metal framework contains several objects:

  • device – an abstraction of the GPU that consumes the rendering and compute commands from the command queue
  • command queue – is a serial sequence of command buffers and establishes the order the stored commands will be executed
  • command buffer – stores the translated commands from a command encoder. Metal notifies the application when the buffer finished executing the commands.
  • command encoder – translates the API commands to GPU hardware commands – there are 3types of encoders: render (for graphics rendering), compute (for data parallel processing) and blit (for resource copy operations). For now we will only look at the render command encoder.
  • states (eg blend and depth)
  • shaders (source code)
  • resources (textures and data buffers)

We will talk about the last 3 objects in the next episode of this series. All we care for now are the devicequeuebuffer and encoder. The Render Command Encoder (RCE) generates hardware commands for a single rendering pass, which means all of the rendering that is sent to a single framebuffer object (set of targets). If another framebuffer (set of targets) needs to be rendered, a new RCE needs to be created. The RCE specifies states for the vertex and fragment stages of the graphics pipeline (which we will talk about in the next episode) and it also interleaves resourcesstate changes and draw calls. A huge advantage of using RCEis the fact that there is no draw-time compilation; the app decides when the compilation and state validation occurs, thus offering the programmers great performance leverage.

Let’s return to our code now. Inside the drawRect(:) method call the render() method:

override func drawRect(dirtyRect: NSRect) {

If you run the app, you will see just a nice solid color like this:

alt text

In the next episode we will finally get to introducing shaders, talk about loading textures and about managing model data. The source code is posted on Github as usual.

Until next time!

2 thoughts on “Using MetalKit part 1”

  1. Small note on this nice article: the MTLCommandQueue creation is rather expensive, so you should not create one instance for every rendered frame. I’d suggest to create a device and command queue at initialization time and cache those objects for later use.

    1. hi Dom. thanks for the comment. you are right, however, this is a 5 years old article that was meant to be a Hello Metal quick and dirty sample code 🙂 in later articles you will see that I am also mentioning that as optimization and best practice.

Leave a Reply

Your email address will not be published. Required fields are marked *