Writing custom shaders

This document explains how to write shaders for pygfx for the WgpuRenderer. This may be useful if you want to improve the existing shaders, add new shaders to pygfx, or if you want to implement custom shaders in your own project.

The shader class

A shader object derives from WorldObjectShader. Its purpose is to provide (templated) shader-code (WGSL), set the corresponding template variables, define what bindings (buffers and textures) are used, and provide details of the pipeline and the rendering.

The shader is associated with a WorldObject-material combination using the register_wgpu_render_function() decorator. This decorator can be applied to your shader class, but it can also be applied to a function that returns multiple shader objects. This can be useful if you want multiple “passes”, like a compute pass to prepare some data.

The shader must implement a few methods. A typical shader is shown below:

from pygfx.renderers.wgpu import (
    register_wgpu_render_function, WorldObjectShader, Binding, RenderMask
)

@register_wgpu_render_function(SomeWorldObject, SomeMaterial)
class SomeShader(WorldObjectShader):

    type = "render"  # must be "render" or "compute"

    def get_bindings(self, wobject, shared):
        # Collect bindings. We must return a dict mapping slot
        # indices to Binding objects. But it's sometimes easier to
        # collect bindings in a list and then convert to a dict.
        bindings = [
            Binding("u_stdinfo", "buffer/uniform", shared.uniform_buffer),
            Binding("u_wobject", "buffer/uniform", wobject.uniform_buffer),
            Binding("u_material", "buffer/uniform", wobject.material.uniform_buffer),
            ...
        ]
        bindings = {i:b for i, b in enumerate(bindings)}
        # Generate the WGSL code for these bindings
        self.define_bindings(0, bindings)
        # The "bindings" are grouped as a dict of dicts. Often only
        # bind-group 0 is used.
        return {
            0: bindings,
        }

    def get_pipeline_info(self, wobject, shared):
        # Result. All fields are mandatory.
        return {
            "primitive_topology": wgpu.PrimitiveTopology.triangle_list,
            "cull_mode": wgpu.CullMode.none,
        }

    def get_render_info(self, wobject, shared):
        n_vertices = ...
        n_instances = 1
        render_mask = wobject.render_mask
        if not render_mask:
            render_mask = RenderMask.all
        # Result. All fields are mandatory. The RenderMask.all is a safe
        # value; other values are optimizations.
        return {
            "indices": (n_vertices, n_instances),
            "render_mask": render_mask,
        }

    def get_code(self):
        # Return combination of code pieces.
        return (
            self.code_definitions()
            + self.code_common()
            + self.code_vertex()
            + self.code_fragment()
        )

    def code_vertex(self):
        return """
        @stage(vertex)
        fn vs_main(@builtin(vertex_index) index: u32) -> @builtin(position) vec4<f32> {
            ...
        }
        """

    def code_fragment(self):
        return """
        @stage(fragment)
        fn fs_main() -> FragmentOutput {
           ...
        }
        """

Remarks:

  • In get_bindings(), the Binding object is used to collect all the required information on a binding.

  • The wgsl code to define a group of bindings can be easily generated using define_bindings().

  • You can also manually define the wgsl code for a binding in cases where this is easier. We recommend using a separate bindgroup for that.

  • By convention, methods that return wgsl code are prefixed with “code”.

  • The render_mask specifies in what passes the object must be drawn. Users can set it on the object, but by default it is “auto” (zero), in which case it must be set by the shader. In the code above it is set to “all” which is a safe option, but if the shader knows that all fragments are opaque or all fragments are transparent, the render_mask can be set accordingly.

Render passes and render_mask

When a scene is rendered, it is likely that it’s not rendered once, but twice: one time for the opaque fragments, and one time for the transparent fragments. This depends on the renderer.blend_mode. It can also be set to just a single (opaque) pass, or a mode that provides improved handling of transparent objects that has more than two passes.

Since the used render targets depend on the blend mode and the render pass, the fragment output is abstracted away for shader authors, as we’ll see further on in this document.

Objects that can have both opaque and transparent fragments, must participate in all render passes. However, objects that only have opaque fragments or only transparent fragments, can be optimized. This is what the render_mask in the previous section is about. In case of doubt RenderMask.all is a safe default.

WGSL code and templating

The shader code is written in WGSL. We use jinja2-templating to allow flexible code generation. Here’s an example:

def get_bindings(self, wobject, shared):
    # Template variables can be set like this
    self["scale"] = 1.2
    ...

def code_vertex(self):
    return """
    @stage(vertex)
    fn vs_main(@builtin(vertex_index) index: u32) -> @builtin(position) vec4<f32> {
        let something = x * {{ scale }};
    }
    """

Note that a change to a templating variable requires a recompilation of the wgpu shader module, which is an expensive operation. Therefore it’s better to use uniforms for things that may change often.

Varyings

Variables passed between vertex shader and fragment shader are called “varyings” in GPU terminology (because they vary as they are interpolated between vertices). In pygfx, each vertex function has a Varyings as output, and this is the input of every fragment function. You don’t have to define the Varyings struct anywhere - pygfx takes care of that based on the attributes that are assigned in the vertex shader. The only catch is that the attributes must be set with an explicit type cast:

def code_vertex(self):
    return """
    @stage(vertex)
    fn vs_main(@builtin(vertex_index) index: u32) -> Varyings {
        ...
        var varyings: Varyings;
        varyings.position = vec4<f32>(screen_pos_ndc, ndc_pos.zw);
        varyings.world_pos = vec3<f32>(world_pos.xyz / world_pos.w);
        return varyings;
    }
    """

def code_fragment(self):
    return """
    @stage(fragment)
    fn fs_main(varyings: Varyings) -> FragmentOutput {
        ...
        let world_pos = varyings.world_pos;
        ...
    }
    """

FragmentOutput

In a somewhat similar way, the output of the fragment shader is predefined. Though in this case the output is determined by the blend mode and render pass (opaque or transparent), and the details are hidden from the shader author. This way, pygfx can support advanced handling of transparency without affecting individual shaders. All fragment functions in pygfx are somewhat like this:

def code_fragment(self):
    return """
    @stage(fragment)
    fn fs_main(varyings: Varyings) -> FragmentOutput {
        ...
        var out = get_fragment_output(varyings.position.z, color);
        return out;
    }
    """

Picking

The output struct of the fragment shader also has a pick field that can be set with pointer picking info. To enable picking for a material, use the pick_write parameter.

cube = gfx.Mesh(
    gfx.box_geometry(200, 200, 200),
    gfx.MeshBasicMaterial(map=tex, opacity=0.8, pick_write=True),
)

The picking info returned can vary based on the shader. For all shaders, it is a u64 into which we can pack as many fields as needed, using the pick_pack() function. The material needs to implement a corresponding _wgpu_get_pick_info() method to unpack the picking info. See e.g. the picking of a mesh:

def code_fragment(self):
    return """
    @stage(fragment)
    fn fs_main(varyings: Varyings) -> FragmentOutput {
        ...
        var out = get_fragment_output(varyings.position.z, color);

        // The builtin write_pick templating variable should be used
        // to ensure picking info is only written in the appropriate render pass
        $$ if write_pick
        // 20 + 26 + 6 + 6 + 6 = 64
        out.pick = (
            pick_pack(varyings.pick_id, 20) +
            pick_pack(varyings.pick_idx, 26) +
            pick_pack(u32(varyings.pick_coords.x * 64.0), 6) +
            pick_pack(u32(varyings.pick_coords.y * 64.0), 6) +
            pick_pack(u32(varyings.pick_coords.z * 64.0), 6)
        );
        $$ endif

        return out;
    }
    """

Clipping planes

For common features that apply to all/most objects, wgsl convenience functions are available. Take clipping planes. One can call apply_clipping_planes() to discard the fragment if it’s outside of the clipping planes. Or use check_clipping_planes() to get a boolean.

def code_fragment(self):
    return """
    @stage(fragment)
    fn fs_main(varyings: Varyings) -> FragmentOutput {
        ...

        apply_clipping_planes(varyings.world_pos);
        var out = get_fragment_output(varyings.position.z, color);
        ...
        return out;
    }
    """

Colormapping

Many materials in pygfx support colormapping. We distinguish between colormaps with image input data, and vertex input data (texture coordinates). The number of channels of the input data must match the dimensionality of the colormap (1D, 2D or 3D).

The base shader class has two corresponding helper functions, and there is a wgsl helper function.

For images / volumes:

def get_bindings(self, wobjwect, shared):
    ...
    extra_bindings = self.define_img_colormap(material.map)
    bindings.extend(extra_bindings)
    ...

def code_fragment(self):
    return """
    @stage(fragment)
    fn fs_main(varyings: Varyings) -> FragmentOutput {
        ...
        let img_value = textureSample(t_img, s_img, texcoord.xy);
        let color = sample_colormap(img_value);
        ...
    }
    """

For points / lines, meshes, etc.:

def get_bindings(self, wobjwect, shared):
    ...
    extra_bindings = self.define_vertex_colormap(material.map, geometry.texcoords)
    bindings.extend(extra_bindings)
    ...

def code_fragment(self):
    return """
    @stage(fragment)
    fn fs_main(varyings: Varyings) -> FragmentOutput {
        ...
        let color = sample_colormap(varyings.texcoord);
        ...
    }
    """

Lights and shadows

TODO

Other functions

Other function that can be used in wgsl are:

  • ndc_to_world_pos(vec4<f32>) -> vec3<f32>