pixijs-custom-rendering

v2026.09.24

Use this skill when writing custom shaders, uniforms, filters, batchers, or low-level draw code in PixiJS v8. Covers Shader.from({gl, gpu, resources}), GlProgram/GpuProgram, UniformGroup with typed uniforms (f32, vec2, mat4x4), UBO mode, textures as resources, TextureView for depth sampling, WGSL override constants, custom bind group layouts (gpuLayout), partial buffer updates (Buffer.update), Geometry.vertexCount/indexCount, TexturePool scratch textures, State winding and culling (clockwiseFrontFace, cullMode), WebGPU render bundles, custom Filter via Filter.from, GLSL ES 3.0 conventions (finalColor), uBackTexture sampling, pixi.js/unsafe-eval for strict CSP, custom Batcher and InstructionPipe.destroyInstructionSet via extensions. Triggers on: Shader, GLSL, WGSL, uniform, custom shader, blendRequired, ShaderOverrides, bind group layout, depth texture, getSize, indexCount, TexturePool, getOptimalTexture, destroyInstructionSet, beginBundle, executeBundle, RenderBundle.

GitHub
安装命令
npx skhub add pixijs/pixijs-custom-rendering
Markdown
SKILL.md

Custom shaders bind GLSL and WGSL programs to scene objects via Shader.from({ gl, gpu, resources }). Uniforms live in typed UniformGroups, textures are passed as separate resources, and the same shader can target both WebGL and WebGPU.

Quick Start

const uniforms = new UniformGroup({
  uTime: { value: 0, type: "f32" },
});

const shader = Shader.from({
  gl: { vertex: vertexSrc, fragment: fragmentSrc },
  resources: { uniforms },
});

const geometry = new MeshGeometry({
  positions: new Float32Array([0, 0, 100, 0, 100, 100, 0, 100]),
  uvs: new Float32Array([0, 0, 1, 0, 1, 1, 0, 1]),
  indices: new Uint32Array([0, 1, 2, 0, 2, 3]),
});

const mesh = new Mesh({ geometry, shader });
app.stage.addChild(mesh);

app.ticker.add(() => {
  shader.resources.uniforms.uniforms.uTime = performance.now() / 1000;
});

Related skills: pixijs-filters (built-in filters), pixijs-scene-mesh (custom geometry), pixijs-performance (batch optimization), pixijs-migration-v8 (shader API migration from v7).

Core Patterns

Dual-renderer shader (WebGL + WebGPU)

import { Shader, GlProgram, GpuProgram, UniformGroup } from "pixi.js";

const glVertex = `...`; // GLSL vertex (write `#version 300 es` yourself if you want WebGL2/GLSL ES 3.0)
const glFragment = `...`; // GLSL fragment
const wgslSource = `...`; // WGSL combined

const shader = Shader.from({
  gl: { vertex: glVertex, fragment: glFragment },
  gpu: {
    // entryPoint names are arbitrary; they must match the @vertex / @fragment
    // function names in your WGSL source. PixiJS ships examples using
    // 'mainVert' / 'mainFrag' but `main` is equally valid.
    vertex: { entryPoint: "mainVert", source: wgslSource },
    fragment: { entryPoint: "mainFrag", source: wgslSource },
  },
  resources: {
    myUniforms: new UniformGroup({
      uColor: { value: new Float32Array([1, 0, 0, 1]), type: "vec4<f32>" },
      uMatrix: { value: new Float32Array(16), type: "mat4x4<f32>" },
    }),
  },
});

If only gl is provided, the shader works with WebGL only. If only gpu is provided, it works with WebGPU only. The compatibleRenderers bitmask is set automatically.

GlProgram does not auto-inject #version 300 es. If you write #version 300 es yourself, PixiJS preserves it and treats the shader as GLSL ES 3.0; otherwise it injects WebGL1 compat macros (#define in varying, #define texture texture2D) and runs the shader as WebGL1-style GLSL. GlProgram always injects a default precision (highp vertex, mediump fragment) and the program name. For GLSL ES 3.0, use in/out instead of attribute/varying, texture() instead of texture2D(), and an out vec4 instead of gl_FragColor.

Textures as resources

Textures are resources, not uniforms. Pass the texture's source and style separately:

import { Shader, UniformGroup, Texture, Assets } from "pixi.js";

const texture = await Assets.load("myImage.png");

const shader = Shader.from({
  gl: { vertex: vertSrc, fragment: fragSrc },
  resources: {
    uTexture: texture.source,
    uSampler: texture.source.style,
    myUniforms: new UniformGroup({
      uAlpha: { value: 1.0, type: "f32" },
    }),
  },
});

// Swap texture at runtime
shader.resources.uTexture = otherTexture.source;

Resources are a flat key-value map. The key must match the uniform/binding name in the shader source.

Resources can also be plain objects (auto-wrapped into UniformGroup):

const shader = Shader.from({
  gl: { vertex: vertSrc, fragment: fragSrc },
  resources: {
    myUniforms: {
      uTime: { value: 0, type: "f32" },
    },
  },
});

UBO mode (Uniform Buffer Objects)

UBO mode packs uniforms into a single GPU buffer. Required for WebGPU; optional (WebGL2+) for WebGL.

import { UniformGroup } from "pixi.js";

const ubo = new UniformGroup(
  {
    uProjection: { value: new Float32Array(16), type: "mat4x4<f32>" },
    uAlpha: { value: 1.0, type: "f32" },
  },
  { ubo: true, isStatic: true },
);

// Must call update() manually when isStatic is true
ubo.uniforms.uAlpha = 0.5;
ubo.update();

UBO rules:

  • Only f32 and i32 based types are supported (no u32). Matrices are float-only.
  • Samplers/textures cannot go in a UBO.
  • The UniformGroup name in resources must exactly match the UBO block name in the shader.
  • Structure and order must exactly match the shader layout.
  • UBO sync uses new Function under the hood. In strict-CSP environments (no unsafe-eval), import pixi.js/unsafe-eval once at startup to swap in the fallback sync path; without it, UBO-backed shaders (and therefore WebGPU) will throw on first use.

Custom filter

Filter.from({ gl, resources }) is the shorthand. Pass only a fragment shader; PixiJS supplies a default vertex shader that handles output frame positioning.

import { Filter } from "pixi.js";

const filter = Filter.from({
  gl: {
    fragment: `
            in vec2 vTextureCoord;
            out vec4 finalColor;
            uniform sampler2D uTexture;
            uniform float uStrength;

            void main(void) {
                vec4 color = texture(uTexture, vTextureCoord);
                finalColor = mix(color, vec4(1.0 - color.rgb, color.a), uStrength);
            }
        `,
  },
  resources: {
    filterUniforms: {
      uStrength: { value: 0.5, type: "f32" },
    },
  },
});

filter.resources.filterUniforms.uniforms.uStrength = 1.0;

For a custom vertex shader, use new Filter({ glProgram: new GlProgram({ vertex, fragment }), resources }).

Filter shader conventions (GLSL ES 3.0)

  • in vec2 vTextureCoord; instead of varying vec2 vTextureCoord;
  • out vec4 finalColor; instead of gl_FragColor
  • texture(uTexture, uv) instead of texture2D(uTexture, uv)
  • The default vertex shader exposes uInputSize, uOutputFrame, uOutputTexture and helpers filterVertexPosition() / filterTextureCoord()

Sampling the render target behind the filter

Set blendRequired: true and sample uBackTexture in the fragment shader. PixiJS copies the destination pixels into that uniform before running the filter:

const blendFilter = Filter.from({
  gl: { fragment: blendFragSrc },
  resources: { uniforms: { uAmount: { value: 0.5, type: "f32" } } },
  blendRequired: true,
});

Only enable blendRequired when you need it; it forces an extra GPU copy every frame.

Updating uniforms at runtime

// Access the UniformGroup via resources
shader.resources.myUniforms.uniforms.uTime = performance.now() / 1000;

// For isStatic UBOs, call update() after changing values
shader.resources.myUniforms.update();

Advanced GPU features

See references/advanced-gpu.md for full samples of:

  • Partial buffer updates: buffer.update(sizeInBytes, offsetInBytes) uploads only the changed byte range.
  • Vertex and index counts, winding, culling: geometry.vertexCount (replaces the deprecated getSize()), geometry.indexCount to draw a prefix of a shared index buffer, state.clockwiseFrontFace and cullMode.
  • WGSL override constants (WebGPU): Shader.from({ gpu, resources, overrides: { STEPS: 8 } }); each distinct set compiles its own pipeline.
  • Custom bind group layouts (WebGPU): generate the default with generateGpuLayoutGroups(extractStructAndGroups(source)), edit it, pass it as gpuLayout.
  • Depth sampling with TextureView (WebGPU): bind new TextureView(depth, { aspect: "depth-only" }) as a resource while the target's depth attachment is depthReadOnly.
  • Render bundles (WebGPU): record draws once with encoder.beginBundle() / endBundle() and replay with executeBundle() while isBundleValid() holds.
  • Pooled scratch textures: TexturePool.getOptimalTexture({ width, height, resolution, antialias }) and returnTexture().

Uniform type reference

See references/uniform-types.md for the complete table of supported types, their WGSL/GLSL equivalents, and value formats.

Custom Batcher (extension-based)

The Batcher abstract class enables custom batching for specialized rendering. Subclass it and register via extensions:

import { Batcher, extensions, ExtensionType } from "pixi.js";
import type {
  BatcherOptions,
  BatchableMeshElement,
  BatchableQuadElement,
  Geometry,
  Shader,
} from "pixi.js";

class MyBatcher extends Batcher {
  public static extension = {
    type: [ExtensionType.Batcher],
    name: "my-batcher",
  };

  public name = "my-batcher";
  protected vertexSize = 6; // floats per vertex
  public geometry: Geometry;
  public shader: Shader;

  constructor(options: BatcherOptions) {
    super(options);
    // Initialize geometry and shader
  }

  public packAttributes(
    element: BatchableMeshElement,
    float32View: Float32Array,
    uint32View: Uint32Array,
    index: number,
    textureId: number,
  ): void {
    // Pack mesh vertex attributes into the batch buffer
  }

  public packQuadAttributes(
    element: BatchableQuadElement,
    float32View: Float32Array,
    uint32View: Uint32Array,
    index: number,
    textureId: number,
  ): void {
    // Pack quad vertex attributes into the batch buffer
  }
}

extensions.add(MyBatcher);

Elements reference the batcher by batcherName. The BatchableElement interface requires: batcherName, texture, blendMode, indexSize, attributeSize, topology, and packAsQuad.

A custom InstructionPipe that caches state per InstructionSet (the way BatcherPipe keeps its batchers) should implement destroyInstructionSet(instructionSet). The renderer calls it when the owning render group is destroyed, so the cached GPU objects are released with it instead of leaking.

public destroyInstructionSet(instructionSet: InstructionSet): void {
  this._cache[instructionSet.uid]?.destroy();
  delete this._cache[instructionSet.uid];
}

Common Mistakes

[CRITICAL] Old Shader.from(vertex, fragment, uniforms) constructor

Wrong:

const shader = Shader.from(vertex, fragment, { uTime: 1 });

Correct:

const shader = Shader.from({
  gl: { vertex, fragment },
  resources: {
    uniforms: new UniformGroup({
      uTime: { value: 1, type: "f32" },
    }),
  },
});

v8 requires an options object with gl/gpu programs and resources. The positional API was removed.

[CRITICAL] UniformGroup without type annotation

Wrong:

new UniformGroup({ uTime: 1 });

Correct:

new UniformGroup({ uTime: { value: 1, type: "f32" } });

Every uniform requires an explicit { value, type } pair. Omitting the type causes a runtime error: "Uniform type undefined is not supported."

[HIGH] UBO with unsupported types or wrong structure

UBO mode supports f32 and i32 based types (scalars and vectors). u32 is not in the supported UniformGroup type list and will throw. Matrices are float-only (mat*<f32>). Samplers cannot be placed in UBOs.

The struct name and field order must exactly match the shader's UBO declaration. Mismatches produce garbled rendering with no error.

[HIGH] Putting textures in UniformGroup

Wrong:

new UniformGroup({
  uTexture: { value: texture, type: "f32" },
});

Correct:

const shader = Shader.from({
  gl: { vertex, fragment },
  resources: {
    uTexture: texture.source,
    uSampler: texture.source.style,
    myUniforms: new UniformGroup({
      uAlpha: { value: 1.0, type: "f32" },
    }),
  },
});

Textures are resources, not uniforms. Pass texture.source (TextureSource) and texture.source.style (TextureStyle) as top-level resource entries.

[HIGH] Replaying a render bundle against a different render target

Wrong:

const bundle = encoder.endBundle();
// later, on every frame, regardless of what is bound
encoder.executeBundle(bundle);

Correct:

if (!bundle || !encoder.isBundleValid(bundle)) {
  bundle = record();
}
encoder.executeBundle(bundle);

A bundle bakes the attachments and winding of the pass it was recorded in, and the device it was recorded on. Replaying it under a different target either fails WebGPU validation for the whole frame or silently renders inside out; replaying it after a device loss is rejected outright.

[MEDIUM] Reading Geometry.getSize()

geometry.getSize() is deprecated since 8.20.0 and logs a warning. Use geometry.vertexCount, which is cached and only recomputed when buffers or attributes change.

API Reference

发现
标签

此技能尚未发布标签。

版本
最新版本元数据

版本

v2026.09.24

发布时间

2026年9月24日

分类

未分类

许可证

MIT

源路径

skills/pixijs-custom-rendering

默认分支

main

最新提交

df555b8

Tree SHA

463d907