gpu.js/scripts/benchmark-webgpu.mjs at develop · gpujs/gpu.js

History

359 lines (339 loc) · 14.4 KB

Raw

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

#!/usr/bin/env node

// Benchmarks cpu vs webgl2 vs webgpu over identical kernel sources against

// dist/gpu-browser.js, in headed Chromium — headless Chromium exposes

// navigator.gpu but requestAdapter() resolves null, so headed is mandatory.

// Prints a GitHub-markdown table plus the raw JSON to stdout (progress goes

// to stderr, nothing is written to disk).

// npm run build && node scripts/benchmark-webgpu.mjs

// Methodology (see scratchpad webgpu/test-bench-plan.md §3):

// - wall clock only; GPU timer queries report garbage on ANGLE Metal

// - every timed run is synced by reading back its own result — gl.finish()

// is not a sync on Metal

// - timed runs ping-pong between two input sets so no driver can elide a

// repeated dispatch (Apple TBDR elides redundant identical draws)

// - median of >= 10 runs, warmup excluded; cpu capped to 3 runs when a

// single run exceeds 2 s

// - every mode's results are cross-checked against cpu (relative 1e-4)

// before timing; a mismatch aborts the whole run

// - webgpu is attempted last: while the backend is absent (or the platform

// has no adapter) it reports "unavailable" and the cpu/webgl2 numbers

// still print; once src/backend/web-gpu lands, the column activates with

// no harness changes (kernel() returns a Promise there — every call site

// awaits, which is a no-op for the sync modes).

import net from 'node:net';

import path from 'node:path';

import { spawn } from 'node:child_process';

import { fileURLToPath } from 'node:url';

import { chromium } from '@playwright/test';

const ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..');

const PORT = 8099;

function portListening(port) {

return new Promise(resolve => {

const socket = net.connect({ port, host: '127.0.0.1' });

socket.once('connect', () => { socket.destroy(); resolve(true); });

socket.once('error', () => resolve(false));

});

}

// Everything below runs inside the page in ONE session: gpu.js parses kernel

// sources via Function.prototype.toString, so the kernels must be defined

// here, not passed in. Must stay closure-free — Playwright serializes it.

async function benchInPage() {

const G = window.GPU;

if (!G || !G.prototype || typeof G.prototype.createKernel !== 'function') {

throw new Error('dist/gpu-browser.js did not expose GPU — run npm run build first');

}

const log = (...a) => console.log(a.join(' '));

// -- deterministic inputs, two sets per workload for ping-ponging --------

function rng(seed) {

let s = seed >>> 0;

return function () {

s = (s + 0x6d2b79f5) >>> 0;

let t = s;

t = Math.imul(t ^ (t >>> 15), t | 1);

t ^= t + Math.imul(t ^ (t >>> 7), t | 61);

return ((t ^ (t >>> 14)) >>> 0) / 4294967296;

};

}

function matrix(n, seed) {

const r = rng(seed);

const m = new Array(n);

for (let y = 0; y < n; y++) {

const row = (m[y] = new Float32Array(n));

for (let x = 0; x < n; x++) row[x] = r();

}

return m;

}

function vector(n, seed) {

const r = rng(seed);

const v = new Float32Array(n);

for (let i = 0; i < n; i++) v[i] = r();

return v;

}

const MAP_N = 4194304;

const inputs = {

mm512: { a: [matrix(512, 1), matrix(512, 3)], b: [matrix(512, 2), matrix(512, 4)] },

mm1024: { a: [matrix(1024, 5), matrix(1024, 7)], b: [matrix(1024, 6), matrix(1024, 8)] },

map: [vector(MAP_N, 9), vector(MAP_N, 10)],

};

// -- kernel sources, identical across all modes --------------------------

function matmulSource(a, b) {

let sum = 0;

for (let i = 0; i < this.constants.size; i++) {

sum += a[this.thread.y][i] * b[i][this.thread.x];

}

return sum;

}

function chain1Source(a) {

return a[this.thread.x] * 2 + 1;

}

function chain2Source(a) {

return a[this.thread.x] * 0.5 + 3;

}

function chain3Source(a) {

return a[this.thread.x] * 1.5 - 2;

}

// the 4M map reuses chain1Source — same source object is fine, each mode

// gets its own kernel instance

// -- helpers -------------------------------------------------------------

function median(sorted) {

const m = sorted.length >> 1;

return sorted.length % 2 ? sorted[m] : (sorted[m - 1] + sorted[m]) / 2;

}

function flatten(result) {

if (result.length && typeof result[0] !== 'number') {

const rows = Array.from(result);

const flat = new Float32Array(rows.length * rows[0].length);

for (let y = 0; y < rows.length; y++) flat.set(rows[y], y * rows[0].length);

return flat;

}

return result;

}

const refs = {};

function gate(key, mode, result) {

const flat = flatten(result);

if (mode === 'cpu') {

refs[key] = flat.slice();

return;

}

const ref = refs[key];

if (!ref) throw new Error('RESULT MISMATCH: no cpu reference for ' + key);

if (flat.length !== ref.length) {

throw new Error('RESULT MISMATCH in ' + key + ' (' + mode + '): length ' + flat.length + ' vs cpu ' + ref.length);

}

let worst = 0, worstI = -1;

for (let i = 0; i < ref.length; i++) {

const d = Math.abs(flat[i] - ref[i]) / Math.max(1, Math.abs(ref[i]));

if (d > worst) { worst = d; worstI = i; }

}

if (worst > 1e-4) {

throw new Error('RESULT MISMATCH in ' + key + ' (' + mode + '): relative error ' + worst +

' at index ' + worstI + ' (got ' + flat[worstI] + ', cpu says ' + ref[worstI] + ')');

}

log(' ', key, mode, 'matches cpu (worst rel err ' + worst.toExponential(2) + ')');

}

// fn(i) must materialize its result (that readback is the sync)

async function timeRuns(fn, { warmup, n, capMs }) {

for (let i = 0; i < warmup; i++) await fn(i);

const times = [];

let target = n;

while (times.length < target) {

const i = times.length;

const t0 = performance.now();

await fn(i);

times.push(performance.now() - t0);

if (capMs && times.length === 1 && times[0] > capMs) target = Math.min(target, 3);

}

times.sort((a, b) => a - b);

return { median: median(times), min: times[0], max: times[times.length - 1], n: times.length, times };

}

// -- one mode, all workloads ---------------------------------------------

async function runMode(mode) {

const gpu = new G({ mode });

const isCpu = mode === 'cpu';

const warmup = isCpu ? 1 : mode === 'webgpu' ? 5 : 3;

const capMs = isCpu ? 2000 : 0;

const out = {};

try {

for (const size of [512, 1024]) {

const key = 'matmul' + size;

const k = gpu.createKernel(matmulSource, { output: [size, size], constants: { size }, precision: 'single' });

const mm = inputs['mm' + size];

const once = async i => await k(mm.a[i % 2], mm.b[i % 2]);

gate(key, mode, await once(0));

out[key] = await timeRuns(once, { warmup, n: 12, capMs });

log(mode, key, out[key].median.toFixed(2), 'ms median of', out[key].n);

}

const mk = gpu.createKernel(chain1Source, { output: [MAP_N], precision: 'single' });

const mapOnce = async i => await mk(inputs.map[i % 2]);

gate('map4m', mode, await mapOnce(0));

out.map4m = await timeRuns(mapOnce, { warmup, n: 12, capMs });

log(mode, 'map4m', out.map4m.median.toFixed(2), 'ms median of', out.map4m.n);

// 3-kernel chain: pipelined modes keep intermediates GPU-resident and

// read back only the final handle

const pipeline = !isCpu;

const kOpts = { output: [MAP_N], precision: 'single', pipeline };

const c1 = gpu.createKernel(chain1Source, kOpts);

const c2 = gpu.createKernel(chain2Source, kOpts);

const c3 = gpu.createKernel(chain3Source, kOpts);

const chainOnce = async i => {

const h1 = await c1(inputs.map[i % 2]);

const h2 = await c2(h1);

const h3 = await c3(h2);

return { out: pipeline ? await h3.toArray() : h3, h3 };

};

gate('chain', mode, (await chainOnce(0)).out);

out.chainIncl = await timeRuns(async i => (await chainOnce(i)).out, { warmup, n: 12, capMs });

log(mode, 'chainIncl', out.chainIncl.median.toFixed(2), 'ms median of', out.chainIncl.n);

if (pipeline) {

// excl-readback is derived: readback of the result is the only

// reliable sync on Metal, so we measure the standalone toArray()

// cost on a settled handle and subtract its median

const { h3 } = await chainOnce(0);

out.readback = await timeRuns(async () => await h3.toArray(), { warmup: 2, n: 10, capMs: 0 });

out.chainExcl = {

median: Math.max(0, out.chainIncl.median - out.readback.median),

derived: 'chainIncl.median - readback.median',

};

log(mode, 'chainExcl', out.chainExcl.median.toFixed(2), 'ms (readback', out.readback.median.toFixed(2), 'ms)');

}

} finally {

const d = gpu.destroy();

if (d && typeof d.then === 'function') await d;

}

return out;

}

// -- environment identification ------------------------------------------

const env = { userAgent: navigator.userAgent, webgl2Renderer: null, webgpuAdapter: null };

try {

const gl = document.createElement('canvas').getContext('webgl2');

const ext = gl.getExtension('WEBGL_debug_renderer_info');

env.webgl2Renderer = gl.getParameter(ext ? ext.UNMASKED_RENDERER_WEBGL : gl.RENDERER);

} catch (e) {

env.webgl2Renderer = 'unavailable: ' + e.message;

}

try {

if (!navigator.gpu) {

env.webgpuAdapter = 'navigator.gpu absent';

} else {

const adapter = await navigator.gpu.requestAdapter();

if (!adapter) {

env.webgpuAdapter = 'no adapter (headless or blocklisted)';

} else {

const info = adapter.info || (adapter.requestAdapterInfo ? await adapter.requestAdapterInfo() : null);

env.webgpuAdapter = info

? [info.vendor, info.architecture, info.device, info.description].filter(Boolean).join(' / ')

: 'adapter present (no info)';

}

} catch (e) {

env.webgpuAdapter = 'error: ' + e.message;

}

// -- the matrix: cpu first (builds references), webgpu last --------------

const modes = {};

modes.cpu = await runMode('cpu');

modes.webgl2 = await runMode('webgl2');

let webgpuStatus;

try {

if (G.isWebGPUSupported === false) {

throw new Error('GPU.isWebGPUSupported is false on this platform');

}

modes.webgpu = await runMode('webgpu');

webgpuStatus = 'ok';

} catch (e) {

const msg = (e && e.message) || String(e);

// wrong answers abort loudly; a missing backend/adapter degrades cleanly

if (msg.indexOf('RESULT MISMATCH') !== -1) throw e;

modes.webgpu = null;

webgpuStatus = 'unavailable: ' + msg;

log('webgpu:', webgpuStatus);

}

return { env, modes, webgpuStatus };

}

// -- report ----------------------------------------------------------------

const ROWS = [

['matmul512', 'matmul 512×512 (incl. readback)'],

['matmul1024', 'matmul 1024×1024 (incl. readback)'],

['map4m', '4M-element map (incl. readback)'],

['chainIncl', '3-kernel pipeline chain, final readback only'],

['chainExcl', '3-kernel pipeline chain, excl. final readback \\*'],

];

function fmtMs(ms) {

return (ms >= 100 ? ms.toFixed(0) : ms >= 10 ? ms.toFixed(1) : ms.toFixed(2)) + ' ms';

}

function printReport(data, chromiumVersion) {

const { env, modes, webgpuStatus } = data;

const lines = [];

lines.push('### Benchmarks — cpu vs webgl2 vs webgpu');

lines.push('');

lines.push('- Chromium ' + chromiumVersion + ' (headed)');

lines.push('- WebGL2 renderer: ' + env.webgl2Renderer);

lines.push('- WebGPU adapter: ' + env.webgpuAdapter);

if (webgpuStatus !== 'ok') lines.push('- webgpu mode: ' + webgpuStatus);

lines.push('- median, warmup excluded; every timed run synced by reading back its own result; inputs ping-ponged between two sets so no dispatch is elidable');

lines.push('');

lines.push('|---|---|---|---|');

for (const [key, label] of ROWS) {

// chainExcl has no cpu leg; its speedup is quoted vs the cpu chain incl readback

const cpuBase = modes.cpu[key === 'chainExcl' ? 'chainIncl' : key];

const cells = ['cpu', 'webgl2', 'webgpu'].map(mode => {

const m = modes[mode];

if (!m) return 'unavailable';

const r = m[key];

if (!r) return '—';

if (mode === 'cpu') return fmtMs(r.median);

const speedup = cpuBase ? ' (' + (cpuBase.median / r.median).toFixed(1) + '× vs cpu)' : '';

return fmtMs(r.median) + speedup;

});

lines.push('| ' + [label, ...cells].join(' | ') + ' |');

}

lines.push('');

lines.push('\\* derived: chain median minus the median cost of a standalone `toArray()` on the settled final handle — readback of the result is the only trustworthy sync on Metal; speedup quoted vs the cpu chain including readback.');

lines.push('');

lines.push('<details><summary>Raw results (JSON)</summary>');

lines.push('');

lines.push('```json');

lines.push(JSON.stringify(data, null, 2));

lines.push('```');

lines.push('');

lines.push('</details>');

process.stdout.write(lines.join('\n') + '\n');

}

// -- driver ----------------------------------------------------------------

async function main() {

let devServer = null;

if (await portListening(PORT)) {

process.stderr.write('dev server already listening on :' + PORT + '\n');

} else {

process.stderr.write('starting dev server on :' + PORT + '\n');

devServer = spawn(process.execPath, [path.join(ROOT, 'scripts', 'dev.js')], {

cwd: ROOT,

env: { ...process.env, PORT: String(PORT) },

stdio: ['ignore', 'ignore', 'inherit'],

});

const deadline = Date.now() + 10000;

while (!(await portListening(PORT))) {

if (Date.now() > deadline) throw new Error('dev server did not start on :' + PORT);

await new Promise(r => setTimeout(r, 200));

}

const browser = await chromium.launch({ headless: false, args: ['--use-angle=metal'] });

try {

const page = await browser.newPage();

page.on('console', msg => process.stderr.write('[page] ' + msg.text() + '\n'));

page.on('pageerror', err => process.stderr.write('[pageerror] ' + err.message + '\n'));

await page.goto('http://localhost:' + PORT + '/test/');

await page.addScriptTag({ url: '/dist/gpu-browser.js' });

const data = await page.evaluate(benchInPage);

printReport(data, browser.version());

} finally {

await browser.close();

if (devServer) devServer.kill();

}

main().catch(err => {

process.stderr.write('\nBENCHMARK ABORTED: ' + (err && err.stack || err) + '\n');

process.exit(1);

});

Navigation Menu

Uh oh!

Uh oh!

FilesExpand file tree

benchmark-webgpu.mjs

Latest commit

History

benchmark-webgpu.mjs

File metadata and controls