Merge pull request #20885 from calixteman/gouraud_gpu

Implement Gouraud-based shading using WebGPU.
This commit is contained in:
calixteman 2026-03-21 15:18:56 +01:00 committed by GitHub
commit 918a319de6
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
7 changed files with 475 additions and 44 deletions

View File

@ -105,6 +105,7 @@ const DefaultPartialEvaluatorOptions = Object.freeze({
iccUrl: null,
standardFontDataUrl: null,
wasmUrl: null,
prepareWebGPU: null,
});
const PatternType = {
@ -1513,7 +1514,8 @@ class PartialEvaluator {
resources,
this._pdfFunctionFactory,
this.globalColorSpaceCache,
localColorSpaceCache
localColorSpaceCache,
this.options.prepareWebGPU
);
patternIR = shadingFill.getIR();
} catch (reason) {

View File

@ -55,7 +55,8 @@ class Pattern {
res,
pdfFunctionFactory,
globalColorSpaceCache,
localColorSpaceCache
localColorSpaceCache,
prepareWebGPU = null
) {
const dict = shading instanceof BaseStream ? shading.dict : shading;
const type = dict.get("ShadingType");
@ -76,6 +77,7 @@ class Pattern {
case ShadingType.LATTICE_FORM_MESH:
case ShadingType.COONS_PATCH_MESH:
case ShadingType.TENSOR_PATCH_MESH:
prepareWebGPU?.();
return new MeshShading(
shading,
xref,
@ -934,7 +936,7 @@ class MeshShading extends BaseShading {
}
_packData() {
let i, ii, j, jj;
let i, ii, j;
const coords = this.coords;
const coordsPacked = new Float32Array(coords.length * 2);
@ -945,25 +947,27 @@ class MeshShading extends BaseShading {
}
this.coords = coordsPacked;
// Stride 4 (RGBA layout, alpha unused) so the buffer maps directly to
// array<u32> in the WebGPU vertex shader without any repacking.
const colors = this.colors;
const colorsPacked = new Uint8Array(colors.length * 3);
const colorsPacked = new Uint8Array(colors.length * 4);
for (i = 0, j = 0, ii = colors.length; i < ii; i++) {
const c = colors[i];
colorsPacked[j++] = c[0];
colorsPacked[j++] = c[1];
colorsPacked[j++] = c[2];
j++; // alpha — unused, stays 0
}
this.colors = colorsPacked;
// Store raw vertex indices (not byte offsets) so the GPU shader can
// address coords / colors without knowing their strides, and so the
// arrays are transferable Uint32Arrays.
const figures = this.figures;
for (i = 0, ii = figures.length; i < ii; i++) {
const figure = figures[i],
ps = figure.coords,
cs = figure.colors;
for (j = 0, jj = ps.length; j < jj; j++) {
ps[j] *= 2;
cs[j] *= 3;
}
const figure = figures[i];
figure.coords = new Uint32Array(figure.coords);
figure.colors = new Uint32Array(figure.colors);
}
}

View File

@ -71,6 +71,20 @@ class BasePdfManager {
FeatureTest.isOffscreenCanvasSupported;
evaluatorOptions.isImageDecoderSupported &&=
FeatureTest.isImageDecoderSupported;
// Set up a one-shot callback so evaluators can notify the main thread that
// WebGPU-acceleratable content was found. The flag ensures the message is
// sent at most once per document.
if (evaluatorOptions.enableWebGPU) {
let prepareWebGPUSent = false;
evaluatorOptions.prepareWebGPU = () => {
if (!prepareWebGPUSent) {
prepareWebGPUSent = true;
handler.send("PrepareWebGPU", null);
}
};
}
delete evaluatorOptions.enableWebGPU;
this.evaluatorOptions = Object.freeze(evaluatorOptions);
// Initialize image-options once per document.

View File

@ -79,6 +79,7 @@ import { DOMFilterFactory } from "./filter_factory.js";
import { DOMStandardFontDataFactory } from "display-standard_fontdata_factory";
import { DOMWasmFactory } from "display-wasm_factory";
import { GlobalWorkerOptions } from "./worker_options.js";
import { initWebGPUMesh } from "./webgpu_mesh.js";
import { Metadata } from "./metadata.js";
import { OptionalContentConfig } from "./optional_content_config.js";
import { PagesMapper } from "./pages_mapper.js";
@ -347,6 +348,7 @@ function getDocument(src = {}) {
? NodeFilterFactory
: DOMFilterFactory);
const enableHWA = src.enableHWA === true;
const enableWebGPU = src.enableWebGPU === true;
const useWasm = src.useWasm !== false;
const pagesMapper = src.pagesMapper || new PagesMapper();
@ -440,6 +442,7 @@ function getDocument(src = {}) {
iccUrl,
standardFontDataUrl,
wasmUrl,
enableWebGPU,
},
};
const transportParams = {
@ -2926,6 +2929,13 @@ class WorkerTransport {
this.#onProgress(data);
});
messageHandler.on("PrepareWebGPU", () => {
if (this.destroyed) {
return;
}
initWebGPUMesh();
});
if (typeof PDFJSDev === "undefined" || !PDFJSDev.test("MOZCENTRAL")) {
messageHandler.on("FetchBinaryData", async data => {
if (this.destroyed) {

View File

@ -13,6 +13,7 @@
* limitations under the License.
*/
import { drawMeshWithGPU, isWebGPUMeshReady } from "./webgpu_mesh.js";
import {
FormatError,
info,
@ -282,7 +283,7 @@ function drawTriangle(data, context, p1, p2, p3, c1, c2, c3) {
const bytes = data.data,
rowSize = data.width * 4;
let tmp;
if (coords[p1 + 1] > coords[p2 + 1]) {
if (coords[p1 * 2 + 1] > coords[p2 * 2 + 1]) {
tmp = p1;
p1 = p2;
p2 = tmp;
@ -290,7 +291,7 @@ function drawTriangle(data, context, p1, p2, p3, c1, c2, c3) {
c1 = c2;
c2 = tmp;
}
if (coords[p2 + 1] > coords[p3 + 1]) {
if (coords[p2 * 2 + 1] > coords[p3 * 2 + 1]) {
tmp = p2;
p2 = p3;
p3 = tmp;
@ -298,7 +299,7 @@ function drawTriangle(data, context, p1, p2, p3, c1, c2, c3) {
c2 = c3;
c3 = tmp;
}
if (coords[p1 + 1] > coords[p2 + 1]) {
if (coords[p1 * 2 + 1] > coords[p2 * 2 + 1]) {
tmp = p1;
p1 = p2;
p2 = tmp;
@ -306,24 +307,24 @@ function drawTriangle(data, context, p1, p2, p3, c1, c2, c3) {
c1 = c2;
c2 = tmp;
}
const x1 = (coords[p1] + context.offsetX) * context.scaleX;
const y1 = (coords[p1 + 1] + context.offsetY) * context.scaleY;
const x2 = (coords[p2] + context.offsetX) * context.scaleX;
const y2 = (coords[p2 + 1] + context.offsetY) * context.scaleY;
const x3 = (coords[p3] + context.offsetX) * context.scaleX;
const y3 = (coords[p3 + 1] + context.offsetY) * context.scaleY;
const x1 = (coords[p1 * 2] + context.offsetX) * context.scaleX;
const y1 = (coords[p1 * 2 + 1] + context.offsetY) * context.scaleY;
const x2 = (coords[p2 * 2] + context.offsetX) * context.scaleX;
const y2 = (coords[p2 * 2 + 1] + context.offsetY) * context.scaleY;
const x3 = (coords[p3 * 2] + context.offsetX) * context.scaleX;
const y3 = (coords[p3 * 2 + 1] + context.offsetY) * context.scaleY;
if (y1 >= y3) {
return;
}
const c1r = colors[c1],
c1g = colors[c1 + 1],
c1b = colors[c1 + 2];
const c2r = colors[c2],
c2g = colors[c2 + 1],
c2b = colors[c2 + 2];
const c3r = colors[c3],
c3g = colors[c3 + 1],
c3b = colors[c3 + 2];
const c1r = colors[c1 * 4],
c1g = colors[c1 * 4 + 1],
c1b = colors[c1 * 4 + 2];
const c2r = colors[c2 * 4],
c2g = colors[c2 * 4 + 1],
c2b = colors[c2 * 4 + 2];
const c3r = colors[c3 * 4],
c3g = colors[c3 * 4 + 1],
c3b = colors[c3 * 4 + 2];
const minY = Math.round(y1),
maxY = Math.round(y3);
@ -494,26 +495,39 @@ class MeshShadingPattern extends BaseShadingPattern {
paddedWidth,
paddedHeight
);
const tmpCtx = tmpCanvas.context;
const data = tmpCtx.createImageData(width, height);
if (backgroundColor) {
const bytes = data.data;
for (let i = 0, ii = bytes.length; i < ii; i += 4) {
bytes[i] = backgroundColor[0];
bytes[i + 1] = backgroundColor[1];
bytes[i + 2] = backgroundColor[2];
bytes[i + 3] = 255;
if (isWebGPUMeshReady()) {
tmpCanvas.context.drawImage(
drawMeshWithGPU(
this._figures,
context,
backgroundColor,
paddedWidth,
paddedHeight,
BORDER_SIZE
),
0,
0
);
} else {
const data = tmpCanvas.context.createImageData(width, height);
if (backgroundColor) {
const bytes = data.data;
for (let i = 0, ii = bytes.length; i < ii; i += 4) {
bytes[i] = backgroundColor[0];
bytes[i + 1] = backgroundColor[1];
bytes[i + 2] = backgroundColor[2];
bytes[i + 3] = 255;
}
}
for (const figure of this._figures) {
drawFigure(data, figure, context);
}
tmpCanvas.context.putImageData(data, BORDER_SIZE, BORDER_SIZE);
}
for (const figure of this._figures) {
drawFigure(data, figure, context);
}
tmpCtx.putImageData(data, BORDER_SIZE, BORDER_SIZE);
const canvas = tmpCanvas.canvas;
return {
canvas,
canvas: tmpCanvas.canvas,
offsetX: offsetX - BORDER_SIZE * scaleX,
offsetY: offsetY - BORDER_SIZE * scaleY,
scaleX,

382
src/display/webgpu_mesh.js Normal file
View File

@ -0,0 +1,382 @@
/* Copyright 2026 Mozilla Foundation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
import { MeshFigureType } from "../shared/util.js";
// WGSL shader for Gouraud-shaded triangle mesh rasterization.
// Vertices arrive in PDF content-space coordinates; the vertex shader
// applies the affine transform supplied via a uniform buffer to map them
// to NDC (X: -1..1 left→right, Y: -1..1 bottom→top).
// Colors are delivered as unorm8x4 (r,g,b,_) and passed through as-is.
const WGSL = /* wgsl */ `
struct Uniforms {
offsetX : f32,
offsetY : f32,
scaleX : f32,
scaleY : f32,
paddedWidth : f32,
paddedHeight : f32,
borderSize : f32,
_pad : f32,
};
@group(0) @binding(0) var<uniform> u : Uniforms;
struct VertexInput {
@location(0) position : vec2<f32>,
@location(1) color : vec4<f32>,
};
struct VertexOutput {
@builtin(position) position : vec4<f32>,
@location(0) color : vec3<f32>,
};
@vertex
fn vs_main(in : VertexInput) -> VertexOutput {
var out : VertexOutput;
let cx = (in.position.x + u.offsetX) * u.scaleX;
let cy = (in.position.y + u.offsetY) * u.scaleY;
out.position = vec4<f32>(
((cx + u.borderSize) / u.paddedWidth) * 2.0 - 1.0,
1.0 - ((cy + u.borderSize) / u.paddedHeight) * 2.0,
0.0,
1.0
);
out.color = in.color.rgb;
return out;
}
@fragment
fn fs_main(in : VertexOutput) -> @location(0) vec4<f32> {
return vec4<f32>(in.color, 1.0);
}
`;
class WebGPUMesh {
#initPromise = null;
#device = null;
#pipeline = null;
// Format chosen to match the OffscreenCanvas swapchain on this device.
#preferredFormat = null;
async #initGPU() {
if (!globalThis.navigator?.gpu) {
return false;
}
try {
const adapter = await navigator.gpu.requestAdapter();
if (!adapter) {
return false;
}
this.#preferredFormat = navigator.gpu.getPreferredCanvasFormat();
const device = (this.#device = await adapter.requestDevice());
const shaderModule = device.createShaderModule({ code: WGSL });
this.#pipeline = device.createRenderPipeline({
layout: "auto",
vertex: {
module: shaderModule,
entryPoint: "vs_main",
buffers: [
{
// Buffer 0: PDF content-space coords, 2 × float32 per vertex.
arrayStride: 2 * 4,
attributes: [
{ shaderLocation: 0, offset: 0, format: "float32x2" },
],
},
{
// Buffer 1: vertex colors, 4 × unorm8 per vertex (r, g, b, _).
arrayStride: 4,
attributes: [
{ shaderLocation: 1, offset: 0, format: "unorm8x4" },
],
},
],
},
fragment: {
module: shaderModule,
entryPoint: "fs_main",
// Use the canvas-preferred format so the OffscreenCanvas swapchain
// and the pipeline output format always agree.
targets: [{ format: this.#preferredFormat }],
},
primitive: { topology: "triangle-list" },
});
return true;
} catch {
return false;
}
}
init() {
if (this.#initPromise === null) {
this.#initPromise = this.#initGPU();
}
}
get isReady() {
return this.#device !== null;
}
/**
* Build flat Float32Array (positions) and Uint8Array (colors) vertex
* streams for non-indexed triangle-list rendering.
*
* Coords and colors intentionally use separate lookup indices. For patch
* mesh figures (types 6/7 converted to LATTICE in the worker), the coord
* index-space and color index-space differ because the stream interleaves
* them at different densities (12 coords but 4 colors per flag-0 patch).
* A single shared index buffer cannot represent both simultaneously, so we
* expand each triangle vertex individually into the two flat streams.
*
* @param {Array} figures
* @param {Object} context coords/colors/offsetX/offsetY/scaleX/scaleY
* @returns {{ posData: Float32Array, colData: Uint8Array,
* vertexCount: number }}
*/
#buildVertexStreams(figures, context) {
const { coords, colors } = context;
// Count vertices first so we can allocate the typed arrays exactly once.
let vertexCount = 0;
for (const figure of figures) {
const ps = figure.coords;
if (figure.type === MeshFigureType.TRIANGLES) {
vertexCount += ps.length;
} else if (figure.type === MeshFigureType.LATTICE) {
const vpr = figure.verticesPerRow;
// 2 triangles × 3 vertices per quad cell
vertexCount += (Math.floor(ps.length / vpr) - 1) * (vpr - 1) * 6;
}
}
// posData: 2 × float32 per vertex (raw PDF content-space x, y).
// colData: 4 × uint8 per vertex (r, g, b, unused — required by unorm8x4).
const posData = new Float32Array(vertexCount * 2);
const colData = new Uint8Array(vertexCount * 4);
let pOff = 0,
cOff = 0;
// pi and ci are raw vertex indices; coords is stride-2, colors stride-4.
const addVertex = (pi, ci) => {
posData[pOff++] = coords[pi * 2];
posData[pOff++] = coords[pi * 2 + 1];
colData[cOff++] = colors[ci * 4];
colData[cOff++] = colors[ci * 4 + 1];
colData[cOff++] = colors[ci * 4 + 2];
cOff++; // alpha channel — unused in the fragment shader
};
for (const figure of figures) {
const ps = figure.coords;
const cs = figure.colors;
if (figure.type === MeshFigureType.TRIANGLES) {
for (let i = 0, ii = ps.length; i < ii; i += 3) {
addVertex(ps[i], cs[i]);
addVertex(ps[i + 1], cs[i + 1]);
addVertex(ps[i + 2], cs[i + 2]);
}
} else if (figure.type === MeshFigureType.LATTICE) {
const vpr = figure.verticesPerRow;
const rows = Math.floor(ps.length / vpr) - 1;
const cols = vpr - 1;
for (let i = 0; i < rows; i++) {
let q = i * vpr;
for (let j = 0; j < cols; j++, q++) {
// Upper-left triangle: q, q+1, q+vpr
addVertex(ps[q], cs[q]);
addVertex(ps[q + 1], cs[q + 1]);
addVertex(ps[q + vpr], cs[q + vpr]);
// Lower-right triangle: q+vpr+1, q+1, q+vpr
addVertex(ps[q + vpr + 1], cs[q + vpr + 1]);
addVertex(ps[q + 1], cs[q + 1]);
addVertex(ps[q + vpr], cs[q + vpr]);
}
}
}
}
return { posData, colData, vertexCount };
}
/**
* Render a mesh shading to an ImageBitmap using WebGPU.
*
* Two flat vertex streams (positions and colors) are uploaded from the
* packed IR typed arrays. A uniform buffer carries the affine transform
* so the vertex shader maps PDF content-space coordinates to NDC without
* any CPU arithmetic per vertex.
*
* After `device.queue.submit()`, `transferToImageBitmap()` presents the
* current GPU frame synchronously the browser ensures all submitted GPU
* commands are complete before returning. The resulting ImageBitmap stays
* GPU-resident; `ctx2d.drawImage(bitmap)` is a zero-copy GPU-to-GPU blit.
*
* The GPU device must already be initialized (`this.isReady === true`).
*
* @param {Array} figures
* @param {Object} context coords/colors/offsetX/offsetY/
* @param {Uint8Array|null} backgroundColor [r,g,b] or null for transparent
* @param {number} paddedWidth render-target width
* @param {number} paddedHeight render-target height
* @param {number} borderSize transparent border size in pixels
* @returns {ImageBitmap}
*/
draw(
figures,
context,
backgroundColor,
paddedWidth,
paddedHeight,
borderSize
) {
const device = this.#device;
const { offsetX, offsetY, scaleX, scaleY } = context;
const { posData, colData, vertexCount } = this.#buildVertexStreams(
figures,
context
);
// Upload vertex positions (raw PDF coords) and colors as separate buffers.
// GPUBufferUsage requires size > 0.
const posBuffer = device.createBuffer({
size: Math.max(posData.byteLength, 4),
usage: GPUBufferUsage.VERTEX | GPUBufferUsage.COPY_DST,
});
if (posData.byteLength > 0) {
device.queue.writeBuffer(posBuffer, 0, posData);
}
const colBuffer = device.createBuffer({
size: Math.max(colData.byteLength, 4),
usage: GPUBufferUsage.VERTEX | GPUBufferUsage.COPY_DST,
});
if (colData.byteLength > 0) {
device.queue.writeBuffer(colBuffer, 0, colData);
}
// Uniform buffer: affine transform parameters for the vertex shader.
const uniformBuffer = device.createBuffer({
size: 8 * 4, // 8 × float32 = 32 bytes
usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,
});
device.queue.writeBuffer(
uniformBuffer,
0,
new Float32Array([
offsetX,
offsetY,
scaleX,
scaleY,
paddedWidth,
paddedHeight,
borderSize,
0, // padding to 32 bytes
])
);
const bindGroup = device.createBindGroup({
layout: this.#pipeline.getBindGroupLayout(0),
entries: [{ binding: 0, resource: { buffer: uniformBuffer } }],
});
// The canvas covers the full padded area so the border is naturally clear.
const offscreen = new OffscreenCanvas(paddedWidth, paddedHeight);
const gpuCtx = offscreen.getContext("webgpu");
gpuCtx.configure({
device,
format: this.#preferredFormat,
// "premultiplied" allows fully transparent border pixels when there is
// no backgroundColor; "opaque" is slightly more efficient otherwise.
alphaMode: backgroundColor ? "opaque" : "premultiplied",
});
const clearValue = backgroundColor
? {
r: backgroundColor[0] / 255,
g: backgroundColor[1] / 255,
b: backgroundColor[2] / 255,
a: 1,
}
: { r: 0, g: 0, b: 0, a: 0 };
const commandEncoder = device.createCommandEncoder();
const renderPass = commandEncoder.beginRenderPass({
colorAttachments: [
{
view: gpuCtx.getCurrentTexture().createView(),
clearValue,
loadOp: "clear",
storeOp: "store",
},
],
});
if (vertexCount > 0) {
renderPass.setPipeline(this.#pipeline);
renderPass.setBindGroup(0, bindGroup);
renderPass.setVertexBuffer(0, posBuffer);
renderPass.setVertexBuffer(1, colBuffer);
renderPass.draw(vertexCount);
}
renderPass.end();
device.queue.submit([commandEncoder.finish()]);
posBuffer.destroy();
colBuffer.destroy();
uniformBuffer.destroy();
// Present the current GPU frame and capture it as an ImageBitmap.
// The browser flushes all pending GPU commands before returning, so this
// is synchronous from the JavaScript perspective. The ImageBitmap is
// GPU-resident; drawing it onto a 2D canvas is a GPU-to-GPU blit.
return offscreen.transferToImageBitmap();
}
}
const _webGPUMesh = new WebGPUMesh();
function initWebGPUMesh() {
_webGPUMesh.init();
}
function isWebGPUMeshReady() {
return _webGPUMesh.isReady;
}
function drawMeshWithGPU(
figures,
context,
backgroundColor,
paddedWidth,
paddedHeight,
borderSize
) {
return _webGPUMesh.draw(
figures,
context,
backgroundColor,
paddedWidth,
paddedHeight,
borderSize
);
}
export { drawMeshWithGPU, initWebGPUMesh, isWebGPUMeshReady };

View File

@ -460,6 +460,11 @@ const defaultOptions = {
value: typeof PDFJSDev !== "undefined" && !PDFJSDev.test("MOZCENTRAL"),
kind: OptionKind.API + OptionKind.PREFERENCE,
},
enableWebGPU: {
/** @type {boolean} */
value: true,
kind: OptionKind.API + OptionKind.PREFERENCE,
},
enableXfa: {
/** @type {boolean} */
value: true,