[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Public WebGL] using the same context with multiple canvases

Here's how I think it should work:

var gl = new WebGLContext();

var db1 = gl.createDrawingBuffer({antialias:true, stencil:false, ...});
var db2 = gl.createDrawingBuffer({antialias:false, stencil:true, ...});

var canvas1 = document.createElement('canvas');
var canvas2 = document.createElement('canvas');

canvas1.drawingBuffer = db1;
canvas2.drawingBuffer = db2;

var draw = function(){

On Wed, Mar 25, 2015 at 4:11 PM, Florian Bösch <pyalot@gmail.com> wrote:
The drawImage idea has some issues:
  • It's currently not hardware accelerated (and I'm not sure it can be).
  • It induces a sync stall because you have a sync-point (drawImage) and in order for that to work, the browser has to run the GL commands up to that point.
  • If you need to render several different resolutions (that's almost always the case) you'll end up resizing the canvases drawing buffer several times each frame.
I don't think it's an adequate workaround. A proper solution is to be able to setup drawing buffers to draw to, and to connect them to a canvas, and whatever you draw while one is bound, is what ends up being displayed in that canvas.

On Wed, Mar 25, 2015 at 4:02 PM, Ashley Gullen <ashley@scirra.com> wrote:
It's possible to render a WebGL-rendered canvas to a 2D context canvas with drawImage(). So it seems to me the workaround is just to have one off-screen WebGL context that does all the real rendering, and then drawImage the result to a 2D canvas in the document. This seems to me to have several advantages:
- works today (nothing more to be specced)
- GPU-accelerated copying (assuming it's possible for drawImage() to be GPU-accelerated when passed a WebGL canvas, regardless of whether real implementations do it today - if they do not then this should be added)
- no need to share resources between contexts or load resources twice, since there is only one context
- can update multiple canvases per frame simply by rendering multiple views in one rAF and copying them out to their respective 2D canvases during rendering

If some canvases are different sizes, then the WebGL context just needs to be as big as the largest. It can change its viewport and add a scissor rectangle to save fillrate during rendering too, and drawImage can specify the source area to render from the source WebGL canvas.

This seems to work OK to me, but perhaps I have missed something? I can see it would be a performance improvement if a single draw call could draw to two canvases at once, perhaps from different views, but the above solution seems workable and AFAICT in theory can stay GPU-side without any readback.

This does not cover rendering from a worker, which I do still think is a valuable addition but appears separate to this particular email thread.


On 25 March 2015 at 13:52, Javi Agenjo <javi.agenjo@gmail.com> wrote:
I just wanted to say that I rely in the same solution as Florian for my materials editor, I considered for a long time using a secondary canvas, but that implies that I must store the Images of every texture (instead of removing them once uploaded to VRAM) in case another Canvas needs those textures, so that increases the RAM I need for the whole application.

I still think that some kind of middle ground solution for those cases would be great, not sharing the same resources but allowing cloning data from one context to another directly, like passing the handler of another texture to the texImage2D even if it those come from the same WebGL Context. (I don't know if anyone has considered that option).

On Wed, Mar 25, 2015 at 8:46 AM, Florian Bösch <pyalot@gmail.com> wrote:
Just yesterday I ran into that problem again (for a PBR material manager) where I'd like to show previews of materials as rendered on a sphere.

Since the main canvas is supersampled and fxaa'ed It'd be expensive to update via readback, so that one is rendered from the context as usual. The previews are also ssaed, but the thumbnails are only 100x100 pixels, so each time a user touches any control (slider, color picker etc.) they're rendered off-screen to an fbo and readback to be put into the thumbnails 2D canvas.

I could probably substitute the thumbnails "off-line" canvas 2D with a second GL canvas for the duration of modification of that material. And while that may work well, it's not a solution to usecases that require more than 1-2 canvases to be updated simultaneously (and neither is the readback and copy in solution).

I run into that sort of problem every couple of months. It's quite annoying.

On Mon, Mar 23, 2015 at 9:29 PM, Aleksandar Rodic <aleksandar.xyz@gmail.com> wrote:
The workaround I came up with was a canvas wrapper which migrates the real canvas and its GL context at render time. The canvases which are not rendering show static stand-in canvas with a copy of the image which was last rendered.

This approach works great if you only update one canvas at a time. Rendering to multiple canvases simultaneously causes expensive canvas migrations (copying image data).


On Mon, Mar 23, 2015 at 11:04 AM Mitchell Williams <mitchellwi@google.com> wrote:
If I understand the issue, it sure would be nice if this worked.

I tried to render the scene from a camera to another window (such as when 3DS Max can view a scene in perspective view, and that scene contains a camera which we can see the scene from the camera's view in another window).

As an alternative, I rendered the scene as a texture map, but all contained within a single canvas.

This would be nice to render a scene to multiple canvases.


On Mon, Mar 23, 2015 at 12:06 AM, Florian Bösch <pyalot@gmail.com> wrote:
The topic of how to draw to multiple canvases from a single context (from either the JS main thread or a worker) has not yet been resolved (correct me if I'm wrong).

A possible resolution seems to be the canvas.setContext method. However, this seems to to be under specified, as just having a WebGL context is not sufficient, you also need a drawing buffer which carries attributes such as stencil, depth, bit-depth, MSAA, etc., and the WebGL 2 specification doesn't seem to contain anything about how to create either a WebGL 2 context or a drawing buffer on its own (It just specifies that a getContext call needs to do both).

Regardless of specification status or respective completeness, no implementation of setContext is available, let alone anything to create WebGL contexts and drawing buffers on their own to be used.

I'll posit that this isn't an acceptable situation. I'll reiterate that quite frequently applications that do provide any kind of UI (like sketchfab, verold, goo, fabric) need to put WebGL driven content into many places inside that UI that isn't a single viewport. That content might find itself CSS transformed, composited, scrolled in/out of view, inside an overflow: scroll container, sandwiched between html elements etc. And that content might need to be updated at each animation frame.

I'm frankly mystified why a topic we've been talking about for well over 3 years and have been promised specifications/solutions for, that is widely recognized as an important issue to solve, is nearly unspecified and completely unsolved.

Mitch Williams

Check out my book "WebGL Hotshot" available at: