wgblas

wgblas is an initiative to implement all the standard level 1, 2, 3 BLAS functions on the top of webgpu.

Available Functions

• sscal
• sswap
• saxpy
• scopy
• sdot
• sasum
• snrm2
• isamax
• srot
• srotm

Requirements

• Node.js 18+

Example usage

This package is not yet published to npm. Clone the repo and install it locally in your project:

git clone https://github.com/manit2004/wgblas.git
cd your-project
npm install /path/to/wgblas
Copy

Example Code Snippet

import { init, cleanup } from "wgblas";
import { sscal } from "wgblas/sscal";
import { randomFloat32Array } from "wgblas/util/random";

const device = await init();

const n = 10;
const alpha = 2.0;
const x = randomFloat32Array(n, -10, 10);

console.log("before:", x);
const result = await sscal(device, n, alpha, x, 1);
console.log("after: ", result);
cleanup();
Copy

Benchmark Mode

Pass { benchmark: true } to init() to enable GPU timestamp queries — BLAS functions will then return { result, gpuTimeMs } instead of just the result.

Note: Here gpuTimeMs is only the gpu compute time which doesn't include device to host and host to device transfer time duration.

import { init, cleanup } from "wgblas";
import { sscal } from "wgblas/sscal";

const device = await init({ benchmark: true });

const n = 5;
const alpha = 2.0;
const x = new Float32Array([1, 2, 3, 4, 5]);

const { result, gpuTimeMs } = await sscal(device, n, alpha, x, 1);
console.log(result);                              // Float32Array [2, 4, 6, 8, 10]
console.log(`compute time: ${gpuTimeMs.toFixed(4)} ms`);

cleanup();
Copy

GpuVector usage

GpuVector keeps data resident on the GPU between operations — upload once, chain any number of operations, read back once. This eliminates the redundant uploads and readbacks between steps, which are often more expensive than the compute itself.

import { init, cleanup } from "wgblas";
import { saxpy } from "wgblas/saxpy";
import { sscal } from "wgblas/sscal";
import { GpuVector } from "wgblas/classes/GpuVector";
import { randomFloat32Array } from "wgblas/util/random";

const device = await init();

const n = 10;
const alpha = 2;
const scale = 0.5;
const x = randomFloat32Array(n, -10, 10);
const y = randomFloat32Array(n, -10, 10);

const xGpu = GpuVector.from(x);
const yGpu = GpuVector.from(y);

console.log("x:      ", x);
console.log("y:      ", y);

// results stay in the GPU.
await saxpy(device, n, alpha, xGpu, 1, yGpu, 1);
await sscal(device, n, scale, yGpu, 1);

// single readback
const result = await yGpu.read();
console.log("result: ", result);

xGpu.destroy();
yGpu.destroy();

cleanup();
Copy