What is node-gpgpu

node-gpgpu is Node.js library for gpu accelerated programming. It allows to write accelerated code using subset of javascript and use it as standard javascript functions.

Dependencies

To install and use node-gpgpu you will need cmake, opencl library and opencl runtime installed.

Installation

npm i node-gpgpu

Build

To build node-gpgpu one has to have opencl installed; after that call npm i and npm run test to verify build.

Examples

One of examples is numerical ingetration on the gpu. More examples can be found in repository node-gpgpu-examples that contains code and links to google collab where you can execute it. You can also find some examples in tests such as test/classes.spec.ts.

import { Gpgpu, KernelContext, Types, kernelEntry, kernelFunction } from 'node-gpgpu';
async function main() {
  const n = 2000;
  const iter = 216;
  const gpgpu = new Gpgpu();

  class PiIntegralKernel extends KernelContext {
    @kernelFunction(Types.number, [Types.number])
    f(x: number) {
      return 2 * this.sqrt(1 - x * x);
    }

    @kernelEntry([
      { type: 'Float32Array', readWrite: 'write' },
      { type: 'Object', readWrite: 'read', shapeObj: { n: Types.number, iter: Types.number } },
    ])
    main(c: Float32Array, opt: { n: number, iter: number }) {
      const id = this.get_global_id(0);

      c[id] = 0.0;
      for (let i = id * opt.iter; i < (id + 1) * opt.iter; i += 1) {
        const dx = 2 / (opt.n * opt.iter);
        const x1 = dx * i - 1;
        const x2 = dx * (i + 1) - 1;

        c[id] += (this.f(x2) + this.f(x1)) * 0.5 * dx;
      }
    }
  }

  const k = gpgpu.createKernel(PiIntegralKernel).setSize([2000], [10]);
  const c = new Float32Array(n);

  await k(c, { n, iter });
  const res = c.reduce((prev, curr) => prev + curr);
  console.log(`Result: ${res}`);
}
main();

Roadmap

• Array attributes: length, pop, push, maxLength, etc.
• Math object: Math.random(), Math.floor(), etc.
• Partial results: allow for getting multiple results from a kernel to limit memory being moved to the kernel.