Skip to content

Using MetalKit part 15

At the end of part 13 we concluded we can make our planet look more realistic in two ways: either apply a texture to it, or add some noise to the planet color. We showed how to add noise in part 14. This week we look at textures and samplers. Textures are useful because they can provide a greater level of detail to surfaces than color computing for each vertex could.

Let’s pick up where we left off in Part 13 since we do not need the noise code this time. First, in MetalView.swift let’s remove the mouseDown function as we are not going to need it anymore. Also remove the mouseBuffer and pos variables, as well as any references to them in the code. Then, create a new texture object:

var texture: MTLTexture!

Next, replace this line (it’s likely you removed it already in the above cleaning step):

commandEncoder.setBuffer(mouseBuffer, offset: 0, atIndex: 2)

with this line:

commandEncoder.setTexture(texture, atIndex: 1)

and also change the buffer index for our timer from 1 to 0:

commandEncoder.setBuffer(timerBuffer, offset: 0, atIndex: 0)

I added an image named texture.jpg in the Resources folder of our playground, but you can add yours instead if you want. Let’s create a function that sets up or texture using this image:

func setUpTexture() {
    let path = NSBundle.mainBundle().pathForResource("texture", ofType: "jpg")
    let textureLoader = MTKTextureLoader(device: device!)
    texture = try! textureLoader.newTextureWithContentsOfURL(NSURL(fileURLWithPath: path!), options: nil)

Next, call this function in our init function:

override public init(frame frameRect: CGRect, device: MTLDevice?) {
    super.init(frame: frameRect, device: device)

Now, let’s clean our kernel in Shaders.metal to only include these lines:

kernel void compute(texture2d<float, access::write> output [[texture(0)]],
                    texture2d<float, access::read> input [[texture(1)]],
                    constant float &timer [[buffer(1)]],
                    uint2 gid [[thread_position_in_grid]])
    float4 color =;
    gid.y = input.get_height() - gid.y;
    output.write(color, gid);

You will first notice that we get the input texture through the [[texture(1)]] attribute since that is the index we set it to in the command encoder. Also, the access we requested for it is read. Then we read it into the color variable, however, it comes in flipped upside-down. In order to fix this, on the next line we just flip the Y coordinate for each pixel. The output image should look like this:

alt text

If you open the image and compare it with the output, you will notice it is now correctly oriented. Next, we want to bring back our planet and the dark sky around it. Replace the output line with this block of code:

int width = input.get_width();
int height = input.get_height();
float2 uv = float2(gid) / float2(width, height);
uv = uv * 2.0 - 1.0;
float radius = 0.5;
float distance = length(uv) - radius;
output.write(distance < 0 ? color : float4(0), gid);

This code looks familiar since we already discussed in the previous chapters how to create the planet and the black space surrounding it. The output image should look like this:

alt text

So far so good! We next want to make our planet rotate again. Replace the output line with this block of code:

uv = fmod(float2(gid) + float2(timer * 100, 0), float2(width, height));
color =;
output.write(distance < 0 ? color : float4(0), gid);

This code again looks familiar from the previous chapters where we discussed how to use timer to animate the planet. The output image should look like this:

alt text

This is a bit awkward! The output looks like a spotlight. Replace the last three lines we added with this block of code:

uv = uv * 2;
radius = 1;
constexpr sampler textureSampler(coord::normalized,
                                 mip_filter::linear );
float3 norm = float3(uv, sqrt(1.0 - dot(uv, uv)));
float pi = 3.14;
float s = atan2( norm.z, norm.x ) / (2 * pi);
float t = asin( norm.y ) / (2 * pi);
t += 0.5;
color = input.sample(textureSampler, float2(s + timer * 0.1, t));
output.write(distance < 0 ? color : float4(0), gid);

First, we scale down to half the size of the texture and set the radius to 1 so we can match the planet object size with the texture size. Then comes the magic. Let me introduce the sampler. A sampler is an object that contains various rendering states that a texture needs to configure: its coordinates, the addressing mode (set to repeat here) and the filtering method (set to linear here). Next, we calculate the normal at each point on the sphere, then we compute the angles around the sphere using the normals. Finally, we calculate the color by sampling it instead of reading it as we did before. There is one more thing to do. In the kernel list of arguments, let’s also reconfigure the texture access to sample instead of read. Replace this line:

texture2d<float, access::read> input [[texture(1)]],

with this line:

texture2d<float, access::sample> input [[texture(1)]],

The output image should look like this:

alt text

Now this is what I call a realistic planet surface! The source code is posted on Github as usual.

Until next time!

Leave a Reply

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