Version: Version 53.0.2785.143 (64-bit)
OS: OS X 10.9.5
What steps will reproduce the problem?
(1) open test.html
I'm trying to diagnose a performance problem from what I think is a pretty simple test case. The same OpenGL calls run fine natively, and profiling didn't really seem to be able to tell me much, so I was hoping someone who knew more about the GPU internals could tell me what might be going on.
Attached is a pretty simple WebGL test that runs very badly for me (MacBook Pro Retina with Intel Iris). I should note that the identical code runs well natively.
It seems like it's related (or at least made worse) by multisampling, you can flip kAntialiasing, it runs smoothly for me without multisampling. On my machine GL_SAMPLES is 4 with multisampling on, so it's not a lot of multisampling.
The performance seems to get bad when the shapes overlap, and then gets horrible (<1 fps) when they go outside the viewport. It's hard to measure what's going on, because it's not reflected in any of the WebGL calls, perhaps it's more something with the compositor, etc? I put a stats.js on top quickly just to show that from the perspective of JavaScript everything is running quickly, however the page basically comes to a crawl and almost kills my laptop (can't command-tab switch apps, etc).
What's happening is pretty simple, there is just a single VBO drawn with GL_TRIANGLES. It's 111 vertices / 37 triangles (vec4 per vertex), so it's not very big at all.
It's drawn 30 times in separate passes, changing just one uniform mat4 per pass. There should be no blending or depth testing.
Timeline doesn't tell me too much, except that the GPU is doing something for 1.4s, it seems to attribute this to uniformMatrix4fv but I think it's probably just that is blocking until the GPU finishes.
I tried looking about about:tracing (some screenshots attached), and again I couldn't really see enough to make sense out of what's going on. It seems there are a lot of kDrawArray entries (many more than the 30 draw calls). You can see with the vsync overlay how many frames are being missed : (
The attached code is a little harder to read than I would have liked because it went through a pre-processor, so for example all of the GL constants became integer constants. Sorry about that, there wasn't an easy way to prevent it. I know maybe the test example seems a bit silly, but it's more of just a reduced reproduction of a performance problem in a realistic usage.
Thanks,
|
Deleted:
Screen Shot 2016-10-21 at 15.05.04.png
187 KB
|
|
|
Deleted:
Screen Shot 2016-10-21 at 15.11.17.png
283 KB
|
|
|
Deleted:
Screen Shot 2016-10-21 at 15.12.15.png
301 KB
|
|
|
Deleted:
Screen Shot 2016-10-21 at 15.12.52.png
337 KB
|
|
|
Deleted:
vsyncs.png
183 KB
|
|
|
Deleted:
test.html
11.9 KB
|
Comment 1 by deanm@chromium.org
, Oct 21 2016