webgl-memory

A library to track webgl-memory

View on GitHub

WebGL-Memory

This is a WebGL-Memory tracker. You add the script to your page before you initialize WebGL and then for a given context you can ask how much WebGL memory you’re using.

Note: This is only a guess as various GPUs have different internal requirements. For example a GPU might require that RGB be expanded internally to RGBA. Similarly a GPU might have alignment requirements. Still, this is likely to give a reasonable approximation.

Usage

<script src="https://greggman.github.io/webgl-memory/webgl-memory.js" crossorigin></script>

or

import 'https://greggman.github.io/webgl-memory/webgl-memory.js';

Then in your code

const ext = gl.getExtension('GMAN_webgl_memory');
...
if (ext) {
  const info = ext.getMemoryInfo();
}

The info returned is

{
  memory: {
    buffer: <bytes used by buffers>
    texture: <bytes used by textures>
    renderbuffer: <bytes used by renderbuffers>
    drawingbuffer: <bytes used by the canvas>
    total: <bytes used in total>
  },
  resources: {
    buffer: <count of buffers>,
    renderbuffer: <count of renderbuffers>
    program: <count of programs>
    query: <count of query objects, WebGL2 only>
    sampler: <count of samplers, WebGL2 only>
    shader: <count of shaders>
    sync: <count of sync objects, WebGL2 only>
    texture: <count of textures>
    transformFeedback: <count of transformfeedbacks, WebGL2 only>
    vertexArray: <count of vertexArrays, only if used or WebGL2>
  }
}

Caveats

  1. You must have WebGL error free code.

    If your code is generating WebGL errors you must fix those first before using this library. Consider using webgl-lint to help find your errors.

  2. Resource reference counting is not supported.

    In WebGL if you delete a WebGLObject (a buffer, a texture, etc..), then, if that object is still attached to something else (a buffer attached to a vertex array, a texture attached to a framebuffer, a shader attached to a program), the object is not actually deleted until it’s detached or the thing it’s attached to is itself deleted
    unless the thing it’s attached to is currently bound. It’s complicated 😭

    Tracking all of that in JavaScript is more work than I was willing to put in ATM. My belief is that the stuff that is still attached is usually not a problem because either (a) you’ll delete the objects that are holding the attachments (b) you’ll detach the attachments by binding new ones (c) you have a leak where you’re creating more and more of these objects the hold attachments in which case you can find the issue by watching your resources counts climb.

    Given that it seemed okay to skip this for now.

  3. texImage2D/3D vs texStorage2D/3D

    Be aware that texImage2D/3D may require double the memory of texStorage2D/3D.

    Based on the design of texImage2D/3D, every mip level can have a different size/format and so until it’s time to draw, there is no way to know if those levels will be updated by the app to be matching. Further, in WebGL2, there’s no way to know before draw time if the app will set TEXTURE_BASE_LEVEL and TEXTURE_MAX_LEVEL to be a texture complete subset of mip levels.

    WebGL-memory does not report this difference because it’s up to the implementation what really happens behind the scenes. In general though, texStorage2D/3D has a much higher probablility of using less memory overall.

    The tradeoff for using texStorage is that the texture’s size is immutable. So, for example, if you wanted to wrap a user’s image to a cube, and then change that image when the user selects a different sized image, with texImage you can just upload the new image to the existing texture. With texStorage you’d be required to create a new texture.

  4. ELEMENT_ARRAY_BUFFER

    Buffers used with ELEMENT_ARRAY_BUFFER may need a second copy in ram. This is because WebGL requires no out of bounds memory access (eg, you have a buffer with 10 vertices but you have an index greater >= 10). This can be handled in 2 ways (1) if the driver advertizes “robustness” then rely on the driver (2) keep a copy of that data in ram and check before draw time that no indices are out of range.

    WebGL-memory does not report this difference because it’s up to the implementation. Further, unlike the texture issue above there is nothing an app can do. Fortunately such buffers are usually a small percent of the data used by most WebGL apps.

Example:

Click here for an Example
Unity example here

Development

git clone https://github.com/greggman/webgl-memory.git
cd webgl-memory
npm install

now serve the folder

npx servez

and go to http://localhost:8080/test?src=true

src=true tells the test harness to use the unrolled source from the src folder where as without it uses webgl-memory.js in the root folder which is built using npm run build.

grep=<some expression> will limit the tests as in ...?src=true&grep=renderbuffer only runs the tests with renderbuffer in their description.

Live Tests

built version
source version

Thoughts

I’m not totally convinced this is the right way to do this. If I was making a webgl app and I wanted to know this stuff I think I’d track it myself by wrapping my own creation functions.

In other words, lets say I wanted to know how many times I call fetch.

const req = await fetch(url);
const text = await req.text();

I’d just refactor that

let fetchCount = 0;
function doFetch(url) {
  fetchCount++;
  return fetch(url);
}

...
const req = await doFetch(url);
const text = await req.text();

No need for some fancy library. Simple.

I could do similar things for WebGL functions.

let textureCount = 0;
function makeTexture(gl) {
  textureCount++;
  return gl.createTexture(gl);
}
function freeTexture(gl, tex) {
  --textureCount;
  gl.deleteTexture(tex);
}

const tex = makeTexture(gl);
...
freeTexture(gl, tex);

Also, even if webgl-memory is an okay way to do it I’m not sure making it an extension was the best way vs just some library you call like webglMemoryTracker.init(someWebGLRenderingContext). I structured it this way just because I used webgl-lint as the basis to get this working.

Licence

MIT