Custom GPU nodes

The Creation Graph is a powerful visual scripting language that can generate shader code through its GPU nodes. Extending this with custom nodes allows for more complex algorithms, custom material types and much more. In this tutorial we will demonstrate how to create some basic GPU nodes. To learn the difference between CPU and GPU nodes, check out Node Types.

A creation graph GPU node needs to be in a .tmsl file, these can be compiled by the shader system. Note that there can only be one creation graph node per .tmsl file, additional definitions will be ignored. If these shaders are placed in the bin/data/shaders/ directory, they will be loaded automatically. .tmsl files are written in a Simplified JSON format with less strict punctuation requirements. For a full reference on the shader files, check out the Shader System Reference.

Cube Node

function: [[
	output.res = x * x * x;

creation_graph_node: {
	name: "tm_cube_node"
	display_name: "Cube"
	category: "Shader/Math"

	inputs: [ 
		{ name: "x" display_name: "X" } 
	outputs: [
		{ name: "res" display_name: "Result" type: { type_of: "x" } }

This node shows you the absolute basics of making a creation graph GPU node. All GPU nodes require two blocks. The function block is where you put the actual shader code. The creation_graph_node is a meta node that defines node I/O and general information.

In this example, the creation_graph_node has several fields, but more can be defined:

  • name must be a unique identifier for the node. It’s a good idea to prefix this with your namespace to make sure it doesn't inadvertently collide with nodes created by other people.
  • display_name is optional and specifies node name to show in the UI. If this is empty, a display name will be generated from the name field.
  • category is an optional path-type string that allows you to group related nodes.
  • inputs is an array of input parameters for the node. A type can be specified for each parameter but it is not required. If you don't specify a type, the type will be generic.
  • outputs is an array of output values for the node.

Note that we didn’t specify a type parameter for our input field. This makes it a fuzzy input and anything that supports the multiplication operator can be passed. Our output parameter does have a type field, but instead of defining a fixed type, it uses a generic syntax that sets the output type to whatever the input type was. For more information about this syntax see the Shader System Reference.

Depth Output Node

Output nodes are more complex than function nodes. Instead of a single function block, these nodes take the form of a render pass that can have variations based on the systems used with it and the connected inputs. The example above creates a very simple material node that displays a gray-scale interpretation of the object’s distance to the viewing camera.

depth_stencil_states: {
	depth_test_enable: true
	depth_write_enable: true
	depth_compare_op: "greater_equal"

raster_states: {
	front_face: "ccw"

imports: [
	{ name: "tm" type: "float4x4" }

vertex_shader: {
	import_system_semantics: [ "vertex_id" ]

	code: [[
		tm_vertex_loader_context ctx;
		float4 vp = load_position(ctx, vertex_id, 0);

		float4 wp = mul(vp, load_tm());
		output.position = mul(wp, load_camera_view_projection());
		return output;

pixel_shader: {
	code: [[
		float2 near_far = load_camera_near_far();
		float depth = linearize_depth(input.position.z, near_far.x, near_far.y) * 0.01f;

		output.buffer0 = float4(linear_to_gamma2(depth), 1); // Base color, alpha
		output.buffer1 = float4(1, 1, 0, 1); // Normal (encoded in signed oct)
		output.buffer2 = float4(0, 0, 0, 1); // Specular, Roughness
		output.velocity = float2(0, 0);
		return output;

The creation_graph_node block for this node is very small. If no outputs are specified, the output will be a Shader Instance. These can be passed to other nodes for rendering, like the Draw Call and Shader Instance output nodes.

creation_graph_node: {
	name: "depth_output"
	display_name: "Depth"
	category: "Shader/Output"

In this example the compile block has the following fields:

  • includes specifies which common shaders this shader is dependent on. In this example, that is the common.tmsl shader because we use the linear_to_gamma2() function from that shader.
  • contexts specifies how this pass should be executed depending on the context. In this example, we only support one context, the viewport. In this context, we want to run during the gbuffer phase so we specify that as our layer. We also want to enable the gbuffer_system as we will be writing to it. Finally we specify that in this context we will enable the gbuffer configuration.
  • configurations are groups of settings. In this example we have one configuration group: gbuffer. This configuration requests three systems, if these systems are not present then we cannot run:
    • The viewer_system is needed to query the camera information.
    • The gbuffer_system allows us to render to the G-Buffer in the opaque pass of the default render pipeline.
    • The vertex_buffer_system allows us to query vertex information from the mesh.
compile: {
	includes: [ "common" ]

	configurations: {
		gbuffer: [{ 
			variations: [{ 
				systems: [ "viewer_system", "gbuffer_system", "vertex_buffer_system" ]

	contexts: {
		viewport: [
			{ layer: "gbuffer" enable_systems: [ "gbuffer_system" ] configuration: "gbuffer" }

Note that the available contexts are defined by the application. Some examples of these in The Machinery editor are viewport, shadow_caster and ray_trace_material.

Note that the layers are defined by the render pipeline used. Some examples from the default render pipeline are: gbuffer, skydome, hdr-transparency, ui.


Simon Renger Simon Renger