Advanced post-processing

Introduction

This tutorial describes an advanced method for post-processing in Godot. In particular, it will explain how to write a post-processing shader that uses the depth buffer. You should already be familiar with post-processing generally and, in particular, with the methods outlined in the custom post-processing tutorial.

In the previous post-processing tutorial, we rendered the scene to a Viewport and then rendered the Viewport in a ViewportContainer to the main scene. One limitation of this method is that we could not access the depth buffer because the depth buffer is only available in spatial shaders and Viewports do not maintain depth information.

Full screen quad

In the custom post-processing tutorial, we covered how to use a Viewport to make custom post-processing effects. There are two main drawbacks of using a Viewport:

  1. The depth buffer cannot be accessed

  2. The effect of the post-processing shader is not visible in the editor

To get around the limitation on using the depth buffer, use a MeshInstance with a QuadMesh primitive. This allows us to use a spatial shader and to access the depth texture of the scene. Next, use a vertex shader to make the quad cover the screen at all times so that the post-processing effect will be applied at all times, including in the editor.

First, create a new MeshInstance and set its mesh to a QuadMesh. This creates a quad centered at position (0, 0, 0) with a width and height of 1. Set the width and height to 2. Right now, the quad occupies a position in world space at the origin; however, we want it to move with the camera so that it always covers the entire screen. To do this, we will bypass the coordinate transforms that translate the vertex positions through the difference coordinate spaces and treat the vertices as if they were already in clip space.

The vertex shader expects coordinates to be output in clip space, which are coordinates ranging from -1 at the left and bottom of the screen to 1 at the top and right of the screen. This is why the QuadMesh needs to have height and width of 2. Godot handles the transform from model to view space to clip space behind the scenes, so we need to nullify the effects of Godot's transformations. We do this by setting the POSITION built-in to our desired position. POSITION bypasses the built-in transformations and sets the vertex position directly.

shader_type spatial;

void vertex() {
  POSITION = vec4(VERTEX, 1.0);
}

Even with this vertex shader, the quad keeps disappearing. This is due to frustum culling, which is done on the CPU. Frustum culling uses the camera matrix and the AABBs of Meshes to determine if the Mesh will be visible before passing it to the GPU. The CPU has no knowledge of what we are doing with the vertices, so it assumes the coordinates specified refer to world positions, not clip space positions, which results in Godot culling the quad when we turn away from the center of the scene. In order to keep the quad from being culled, there are a few options:

  1. Add the QuadMesh as a child to the camera, so the camera is always pointed at it

  2. Set the Geometry property extra_cull_margin as large as possible in the QuadMesh

The second option ensures that the quad is visible in the editor, while the first option guarantees that it will still be visible even if the camera moves outside the cull margin. You can also use both options.

Depth texture

To read from the depth texture, perform a texture lookup using texture() and the uniform variable DEPTH_TEXTURE.

float depth = texture(DEPTH_TEXTURE, SCREEN_UV).x;

Note

Similar to accessing the screen texture, accessing the depth texture is only possible when reading from the current viewport. The depth texture cannot be accessed from another viewport to which you have rendered.

The values returned by DEPTH_TEXTURE are between 0 and 1 and are nonlinear. When displaying depth directly from the DEPTH_TEXTURE, everything will look almost white unless it is very close. This is because the depth buffer stores objects closer to the camera using more bits than those further, so most of the detail in depth buffer is found close to the camera. In order to make the depth value align with world or model coordinates, we need to linearize the value. When we apply the projection matrix to the vertex position, the z value is made nonlinear, so to linearize it, we multiply it by the inverse of the projection matrix, which in Godot, is accessible with the variable INV_PROJECTION_MATRIX.

Firstly, take the screen space coordinates and transform them into normalized device coordinates (NDC). NDC run from -1 to 1, similar to clip space coordinates. Reconstruct the NDC using SCREEN_UV for the x and y axis, and the depth value for z.

void fragment() {
  float depth = texture(DEPTH_TEXTURE, SCREEN_UV).x;
  vec3 ndc = vec3(SCREEN_UV, depth) * 2.0 - 1.0;
}

Convert NDC to view space by multiplying the NDC by INV_PROJECTION_MATRIX. Recall that view space gives positions relative to the camera, so the z value will give us the distance to the point.

void fragment() {
  ...
  vec4 view = INV_PROJECTION_MATRIX * vec4(ndc, 1.0);
  view.xyz /= view.w;
  float linear_depth = -view.z;
}

Because the camera is facing the negative z direction, the position will have a negative z value. In order to get a usable depth value, we have to negate view.z.

The world position can be constructed from the depth buffer using the following code. Note that the CAMERA_MATRIX is needed to transform the position from view space into world space, so it needs to be passed to the fragment shader with a varying.

varying mat4 CAMERA;

void vertex() {
  CAMERA = CAMERA_MATRIX;
}

void fragment() {
  ...
  vec4 world = CAMERA * INV_PROJECTION_MATRIX * vec4(ndc, 1.0);
  vec3 world_position = world.xyz / world.w;
}

An optimization

You can benefit from using a single large triangle rather than using a full screen quad. The reason for this is explained here. However, the benefit is quite small and only beneficial when running especially complex fragment shaders.

Set the Mesh in the MeshInstance to an ArrayMesh. An ArrayMesh is a tool that allows you to easily construct a Mesh from Arrays for vertices, normals, colors, etc.

Now, attach a script to the MeshInstance and use the following code:

extends MeshInstance

func _ready():
  # Create a single triangle out of vertices:
  var verts = PoolVector3Array()
  verts.append(Vector3(-1.0, -1.0, 0.0))
  verts.append(Vector3(-1.0, 3.0, 0.0))
  verts.append(Vector3(3.0, -1.0, 0.0))

  # Create an array of arrays.
  # This could contain normals, colors, UVs, etc.
  var mesh_array = []
  mesh_array.resize(Mesh.ARRAY_MAX) #required size for ArrayMesh Array
  mesh_array[Mesh.ARRAY_VERTEX] = verts #position of vertex array in ArrayMesh Array

  # Create mesh from mesh_array:
  mesh.add_surface_from_arrays(Mesh.PRIMITIVE_TRIANGLES, mesh_array)

Note

The triangle is specified in normalized device coordinates. Recall, NDC run from -1 to 1 in both the x and y directions. This makes the screen 2 units wide and 2 units tall. In order to cover the entire screen with a single triangle, use a triangle that is 4 units wide and 4 units tall, double its height and width.

Assign the same vertex shader from above and everything should look exactly the same.

The one drawback to using an ArrayMesh over using a QuadMesh is that the ArrayMesh is not visible in the editor because the triangle is not constructed until the scene is run. To get around that, construct a single triangle Mesh in a modelling program and use that in the MeshInstance instead.