README.md

Canvas CPU benchmark

How fast can we write to a canvas, pixel by pixel, using solely the CPU (i.e., no WebGL)?

This benchmark focus on writing to a buffer and handing it to the canvas context. Some variations are attempted and measured to understand how to efficiently paint the canvas for tasks that require pixel manipulation.

Writing four bytes at once

Instead of writing every pixel one channel at a time:

for (let i = 0; i < canvasSizeInBytes; i += 4) {
    buffer[i] = 0;
    buffer[i+1] = 255;
    buffer[i+2] = 0;
    buffer[i+3] = 255;
}

It is more efficient to write once per pixel using Uint32Arrays:

const view = new Uint32Array(buffer.buffer);

for (let i = 0; i < canvasSizeInPixels; i++) {
    view[i] = (255 << 24) & (255 << 8);
}

This has a huge impact on performance, since gains scale with the number of pixels.

Avoid calling getImageData()

This is a minor, since it'll only happen once per frame, but anyway. Instead of getting a new buffer for every new frame draw:

const imageData = ctx.getImageData(0, 0, width, height);
const buffer = imageData.data;

The best way is to create a single buffer once and reuse it during execution. Just move it outside the frame rendering function - it will work just the same.

Some results

System:

• OS: Intel® Core™ i5-4670 CPU @ 3.40GHz × 4
• CPU: Intel® Core™ i5-4670 CPU @ 3.40GHz × 4

Each test ran for 2 minutes. Results:

• canvas size: 500 x 500
• Uint8ClampedArray: 7.4ns per pixel
• Uint32Array: 4.7ns per pixel

So the speedup was of 1.57.

But time per pixel curiously decreases when the canvas gets bigger:

• canvas size: 1000 x 1000
• Uint8ClampedArray: 4.5ns per pixel
• Uint32Array: 2.9ns per pixel

Speedup still similar, though: 1.55.