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

History

202 lines (188 loc) · 6.74 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

#!/usr/bin/env node

// Benchmarks the webasm backend against cpu and headlessgl in plain Node —

// no browser needed, which is itself the point of this backend. Prints a

// GitHub-markdown table plus raw JSON to stdout.

// node scripts/benchmark-webasm.mjs

// Methodology (matches scripts/benchmark-webgpu.mjs):

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

// before timing; a mismatch aborts the run

// - timed runs ping-pong between two input sets so no cache or memoization

// can elide repeated work

// - median of >= 9 runs, warmup excluded; cpu capped to 3 runs when a

// single run exceeds 2 s

// - webasm rows: scalar (run_simd disabled through the same dispatch the

// kernel uses), SIMD (the sync default), and threaded+SIMD (asyncMode,

// the worker pool -- each worker runs the same run_simd export, so this

// row is both axes compounded; result copied out of shared memory like

// any real caller)

// - the divergent workload exists to price mask predication honestly:

// both branch sides execute for every lane

import { createRequire } from 'node:module';

const require = createRequire(import.meta.url);

const { GPU } = require('../src');

const MEDIAN_RUNS = 9;

function median(times) {

const sorted = [...times].sort((a, b) => a - b);

return sorted[Math.floor(sorted.length / 2)];

}

function relativeError(a, b) {

let worst = 0;

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

const denominator = Math.max(Math.abs(a[i]), Math.abs(b[i]), 1e-20);

worst = Math.max(worst, Math.abs(a[i] - b[i]) / denominator);

}

return worst;

}

function flatten(result) {

if (result[0] && result[0].length !== undefined) {

const out = [];

for (const row of result) out.push(...row);

return out;

}

return Array.from(result);

}

const SIZE = 512;

function makeMatrix(seed) {

const m = [];

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

const row = new Float32Array(SIZE);

for (let x = 0; x < SIZE; x++) row[x] = ((x * 31 + y * 17 + seed) % 100) / 100;

m.push(row);

}

return m;

}

const MAP_N = 4 * 1024 * 1024;

function makeVector(seed, n) {

const v = new Float32Array(n);

for (let i = 0; i < n; i++) v[i] = ((i * 13 + seed) % 1000) / 500 - 1;

return v;

}

const WORKLOADS = [

{

source: function (a, b) {

let sum = 0;

for (let i = 0; i < 512; i++) {

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

}

return sum;

output: [SIZE, SIZE],

inputs: [[makeMatrix(1), makeMatrix(2)], [makeMatrix(3), makeMatrix(4)]],

{

source: function (v) {

const x = v[this.thread.x];

return x * x * 0.5 + Math.sqrt(Math.abs(x)) - x * 0.25;

output: [MAP_N],

inputs: [[makeVector(1, MAP_N)], [makeVector(2, MAP_N)]],

{

source: function (v) {

// three lane-dependent branches plus a lane-varying trip count: the

// shape mask predication exists for

const x = v[this.thread.x];

let acc = 0;

if (x > 0.5) {

acc = x * x * 3;

} else if (x > 0) {

acc = Math.sqrt(x) * 2;

} else {

acc = -x;

}

for (let i = 0; i < (this.thread.x % 7) + 1; i++) {

acc += 0.125;

}

return acc;

output: [1024 * 1024],

inputs: [[makeVector(5, 1024 * 1024)], [makeVector(6, 1024 * 1024)]],

loopMaxIterations: 16,

];

async function timeMode(workload, label, makeKernel, runKernel) {

const kernel = makeKernel();

// correctness gate against cpu before any timing

const cpuGpu = new GPU({ mode: 'cpu' });

const cpuKernel = cpuGpu.createKernel(workload.source, {

output: workload.output,

loopMaxIterations: workload.loopMaxIterations || 1000,

});

const expected = flatten(cpuKernel.apply(null, workload.inputs[0]));

const actual = flatten(await runKernel(kernel, workload.inputs[0]));

const err = relativeError(expected, actual);

if (err > 1e-4) {

throw new Error(`RESULT MISMATCH in ${ workload.name } (${ label }): relative error ${ err }`);

}

cpuGpu.destroy();

await runKernel(kernel, workload.inputs[1]); // warmup second shape

const times = [];

const runs = label === 'cpu' ? 3 : MEDIAN_RUNS;

for (let i = 0; i < runs; i++) {

const inputs = workload.inputs[i % 2];

const start = process.hrtime.bigint();

await runKernel(kernel, inputs);

times.push(Number(process.hrtime.bigint() - start) / 1e6);

}

return { label, ms: +median(times).toFixed(2), err };

}

async function main() {

const rows = [];

for (const workload of WORKLOADS) {

const row = { name: workload.name, modes: {} };

const settings = { output: workload.output, loopMaxIterations: workload.loopMaxIterations || 1000 };

const gpus = [];

const make = (mode, extra) => {

const gpu = new GPU({ mode });

gpus.push(gpu);

return gpu.createKernel(workload.source, Object.assign({}, settings, extra));

};

const configs = [

['cpu', () => make('cpu'), (k, i) => k.apply(null, i)],

['headlessgl', () => make('headlessgl'), (k, i) => k.apply(null, i)],

['webasm scalar', () => {

const kernel = make('webasm');

kernel.apply(null, workload.inputs[0]); // build, then disable simd

kernel.kernel._active.runSimd = null;

return kernel;

}, (k, i) => k.apply(null, i)],

['webasm SIMD', () => make('webasm'), (k, i) => k.apply(null, i)],

['webasm threaded+SIMD', () => make('webasm', { asyncMode: true }), (k, i) => k.apply(null, i)],

];

for (const [label, makeKernel, runKernel] of configs) {

try {

const result = await timeMode(workload, label, makeKernel, runKernel);

row.modes[label] = result.ms;

process.stderr.write(`${ workload.name } / ${ label }: ${ result.ms } ms (err ${ result.err.toExponential(1) })\n`);

} catch (error) {

row.modes[label] = null;

process.stderr.write(`${ workload.name } / ${ label }: FAILED ${ error.message }\n`);

throw error;

}

for (const gpu of gpus) await gpu.destroy();

rows.push(row);

}

const labels = ['cpu', 'headlessgl', 'webasm scalar', 'webasm SIMD', 'webasm threaded+SIMD'];

console.log(`\n| Workload | ${ labels.join(' | ') } |`);

console.log(`|---|${ labels.map(() => '---').join('|') }|`);

for (const row of rows) {

const cpuMs = row.modes['cpu'];

console.log(`| ${ row.name } | ${ labels.map(label => {

const ms = row.modes[label];

if (ms === null) return 'n/a';

const speedup = label === 'cpu' ? '' : ` (${ (cpuMs / ms).toFixed(1) }×)`;

return `${ ms } ms${ speedup }`;

}).join(' | ') } |`);

}

console.log('\n' + JSON.stringify(rows, null, 2));

}

main().catch(error => {

console.error(error);

process.exit(1);

});

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

Uh oh!

Expand file tree

Breadcrumbs

benchmark-webasm.mjs

Latest commit

History

Breadcrumbs

benchmark-webasm.mjs

File metadata and controls