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

Re: [Public WebGL] Problem with 'more-than-65536-points.html' test



This sounds like a driver bug; multisampling should not cause cracks to appear between adjacent triangles. However, I agree with you that the test should focus specifically on rendering a large number of triangles. I've committed an update which disables antialiasing for this test, since it was a much larger refactoring to turn it into a reference test.

-Ken

On Mon, Sep 26, 2011 at 7:51 PM, Jeff Gilbert <jgilbert@mozilla.com> wrote:
It also appears to exhibit the same behavior on linux.
It may very well be an issue with the nVidia drivers for my GPU, but my point about the aim of the test stands.

-Jeff

----- Original Message -----
From: "Jeff Gilbert" <jgilbert@mozilla.com>
To: "Gregg Tavares (wrk)" <gman@google.com>
Cc: "Kenneth Russell" <kbr@google.com>, "public webgl" <public_webgl@khronos.org>
Sent: Monday, September 26, 2011 6:19:37 PM
Subject: Re: [Public WebGL] Problem with 'more-than-65536-points.html' test


"I don't see how this test would have anything to do with anti-aliasing. "
Exactly. It is confusing that this test should fail because of anti-aliasing. Right now, I am getting false-positives, where the test is failing, but what is being tested by the test is working properly.

The issue lies in that it doesn't draw actual quads; it draws pairs of tris, which should be the same thing as quads. However, this is evidently not the case on one of my rendering paths, at least with 2xAA. (0x and 8x work fine)

Regardless, this is a problem that seems to only occur for some sampling patterns. Making the test single-sampled will make it more reliable by further minimizing false-positive errors.

Really what this test should do is first run using only 2 quads (one red, second one green), then run using 65536 / 6 + 2 quads (n-1 red, last one green), and compare the results. If the buffers are the same, then there is no bug in drawing large numbers of points, whether or not there are other problems with rendering.

To clarify, the buffer I'm seeing on affected paths is green, with the exception of a faint red line along part of the seam between the two tris. (seam is bottom-left to top-right)

-Jeff

----- Original Message -----
From: "Gregg Tavares (wrk)" <gman@google.com>
To: "Jeff Gilbert" <jgilbert@mozilla.com>
Cc: "Kenneth Russell" <kbr@google.com>, "public webgl" <public_webgl@khronos.org>
Sent: Monday, September 26, 2011 5:47:14 PM
Subject: Re: [Public WebGL] Problem with 'more-than-65536-points.html' test

I don't see how this test would have anything to do with anti-aliasing.


It just draws overlapping quads. It draws overlapping quads, all at the same location, the only difference being the last quad is a green instead of red.


All the test does is check the last quad gets drawn by checking for green


(in other words, if there was a bug that limited count to 65535 then the last quad would not be drawn and you'd see red)









On Mon, Sep 26, 2011 at 4:38 PM, Jeff Gilbert < jgilbert@mozilla.com > wrote:



Yes, it's bleed-through from between the two tris, a line of faint reddish pixels along part of the seam. This is the only test which is regressing from without antialiasing.
There's clearly a problem somewhere, but it isn't a problem with the number of points: When I reduce the number of quads to only two (one red, then one green), the problem persists.

If we want to test for bleed-through with AA, we should absolutely do so, but (if possible) not in an unrelated test. I'm willing to write up a test if we want to add this.

I agree that it's odd that this is the only test where this is apparently a problem, however I would like to be sure that when this test fails, that it's not just because of antialiasing issues, and is instead because of improper handling of large numbers of points.

-Jeff





----- Original Message -----
From: "Kenneth Russell" < kbr@google.com >
To: "Jeff Gilbert" < jgilbert@mozilla.com >
Cc: "public webgl" < public_webgl@khronos.org >
Sent: Monday, September 26, 2011 4:12:24 PM
Subject: Re: [Public WebGL] Problem with 'more-than-65536-points.html' test

How exactly is the test failing? Are non-green pixels bleeding through
at the edges of the canvas, or is something else happening?

I don't think we've seen any failures of this test on any platform in
Chromium, where antialiasing support for WebGL has been present for
some time.

It's strange that only this test would be affected by the presence of
antialiasing. Are you sure that others aren't affected?

In principle it would be fine with me to disable antialiasing for this
test, but I feel that would probably be a hack which doesn't really
solve the problem.

-Ken

On Mon, Sep 26, 2011 at 3:04 PM, Jeff Gilbert < jgilbert@mozilla.com > wrote:
>
> Hi,
>
> In implementing antialiasing for Firefox, I am getting failures on this test for some rendering paths, and only for some sampling levels.
> While it's possible that this is a driver bug (since this is only on our native GL path, and not with ANGLE), it doesn't seem to me that this test should be using antialiasing. A problem with antialiasing shouldn't fail it even though the issue it's supposed to be testing passes.
>
> Just disabling antialiasing in this test would keep this test specific. If we want to test for the antialiasing artifacts that are causing this test to fail, I think we should make that a separate test.
>
>
> -Jeff
>
> -----------------------------------------------------------
> You are currently subscribed to public_webgl@khronos.org .
> To unsubscribe, send an email to majordomo@khronos.org with
> the following command in the body of your email:
> unsubscribe public_webgl
> -----------------------------------------------------------
>
>

-----------------------------------------------------------
You are currently subscribed to public_webgl@khronos.org .
To unsubscribe, send an email to majordomo@khronos.org with
the following command in the body of your email:
unsubscribe public_webgl
-----------------------------------------------------------



-----------------------------------------------------------
You are currently subscribed to public_webgl@khronos.org.
To unsubscribe, send an email to majordomo@khronos.org with
the following command in the body of your email:
unsubscribe public_webgl
-----------------------------------------------------------