Bug 44459 – fast/canvas/webgl/tex-image-and-sub-image-2d-with-video.html failing on Snow Leopard bot

• Status: RESOLVED
• Resolution: WONTFIX
• Component: WebGL

[Kenneth Russell]

See:

http://build.webkit.org/console
http://build.webkit.org/builders/SnowLeopard%20Intel%20Release%20(Tests)
http://build.webkit.org/builders/SnowLeopard%20Intel%20Release%20(Tests)/builds/16143
http://build.webkit.org/results/SnowLeopard%20Intel%20Release%20(Tests)/r65832%20(16143)/
http://build.webkit.org/results/SnowLeopard%20Intel%20Release%20(Tests)/r65832%20(16143)/results.html
http://build.webkit.org/results/SnowLeopard%20Intel%20Release%20(Tests)/r65832%20(16143)/fast/canvas/webgl/tex-image-and-sub-image-2d-with-video-pretty-diff.html

It looks like random data is getting uploaded to OpenGL, probably because the source movie is in an unsupported format. This fails similarly on my desktop Snow Leopard workstation.

[Kenneth Russell]

Note: this is failing at least with WebKit (run-webkit-tests --debug LayoutTests/fast/canvas/webgl).

[Kenneth Russell]

Committed r65955: <http://trac.webkit.org/changeset/65955>

[Kenneth Russell]

Oops, I unintentionally closed this bug with webkit-patch land. Checked in a disabling of this test on the bots to get them green again. See LayoutTests/platform/mac-snowleopard/Skipped and the above commit, which will need to be reverted when submitting the official fix.

[Eric Seidel (no email)]

many webkit-patch commands respect --no-close which you may find useful in the future.  see webkit-patch help land for more information.

[Adrienne Walker]

This original bug appears to be due to a poorly generated mp4 file.  The theora video has the right colors, but the mp4 has slightly incorrect colors.

After discussion with gman, I decided to just convert the existing test from the Khronos conformance test which uses a red-green video like all of the other tex-image tests for consistency.

Unfortunately, that test failed in both Chromium and Safari on 10.6.

After some investigation, this bug appears to be due to something in Cg, rather than something video-related.  The video appears fine in a <video> tag.  The <video> draws correctly into a 2D canvas.  However, using the <video> or the <canvas> (with the video drawn into it) as an input to texImage2D in WebGL results in black pixels.

Oddly, this only occurs with the red-green video and not with the original video from this bug.

No solution yet, but I just wanted to leave a comment about its progress.

[Adrienne Walker]

Here's the link to the test case that will get checked in once it's working: https://cvs.khronos.org/svn/repos/registry/trunk/public/webgl/sdk/tests/conformance/tex-image-and-sub-image-2d-with-video.html

After some discussion with dhyatt on irc, he suggested looking at the ImageBufferCG data after the call to HTMLVideoElement::paintCurrentFrameIntoContext.  This turned out to all be [0,0,0,255], which leads me back to believe that maybe there is something wrong on the video end.

However, I still don't understand why painting the current frame of a video into an ImageBuffer works in one case (2D canvas) but doesn't in this case (WebGL canvas).

CCing Eric Carlson on this bug, who might have some thoughts about why this is occurring.

[Adrienne Walker]

Created attachment 78611 [details]
Patch

[Adrienne Walker]

(In reply to comment #7)
> Created an attachment (id=78611) [details]
> Patch

Whatever bug was causing this to fail on OSX with the video that was checked into Khronos appears to have been fixed.

[Kenneth Russell]

Comment on attachment 78611 [details]
Patch

Looks OK. Please watch the bots after committing this in case something unexpected happens.

[Zhenyao Mo]

Is webm also supported in Safari?

[Adrienne Walker]

Committed r75570: <http://trac.webkit.org/changeset/75570>

[Adrienne Walker]

Reopened again, argh.

Reverted expectations here: http://trac.webkit.org/changeset/75579

Layout test errors here: http://build.webkit.org/results/SnowLeopard%20Intel%20Release%20(Tests)/r75570%20(23470)/fast/canvas/webgl/tex-image-and-sub-image-2d-with-video-pretty-diff.html

[Adrienne Walker]

(In reply to comment #12)
> Reopened again, argh.
> 
> Reverted expectations here: http://trac.webkit.org/changeset/75579
> 
> Layout test errors here: http://build.webkit.org/results/SnowLeopard%20Intel%20Release%20(Tests)/r75570%20(23470)/fast/canvas/webgl/tex-image-and-sub-image-2d-with-video-pretty-diff.html

This is awfully curious.  If I run the exact command line from http://build.webkit.org/builders/SnowLeopard%20Intel%20Release%20(Tests)/builds/23470/steps/layout-test/logs/stdio on my Intel Snow Leopard box, this test reliably passes.

[Kenneth Russell]

(In reply to comment #13)
> (In reply to comment #12)
> > Reopened again, argh.
> > 
> > Reverted expectations here: http://trac.webkit.org/changeset/75579
> > 
> > Layout test errors here: http://build.webkit.org/results/SnowLeopard%20Intel%20Release%20(Tests)/r75570%20(23470)/fast/canvas/webgl/tex-image-and-sub-image-2d-with-video-pretty-diff.html
> 
> This is awfully curious.  If I run the exact command line from http://build.webkit.org/builders/SnowLeopard%20Intel%20Release%20(Tests)/builds/23470/steps/layout-test/logs/stdio on my Intel Snow Leopard box, this test reliably passes.

Could it have something to do with whether the machine has a monitor attached? What happens if you reboot your Snow Leopard machine with no monitor attached, ssh in and run the test?

[Adrienne Walker]

(In reply to comment #14)

> Could it have something to do with whether the machine has a monitor attached? What happens if you reboot your Snow Leopard machine with no monitor attached, ssh in and run the test?

All the webgl Layout tests crash if I do that.

I guess I don't follow how not having a monitor attached would affect this test.   Isn't DumpRenderTree pushing everything through Mesa?

[Zhenyao Mo]

(In reply to comment #15)
> (In reply to comment #14)
> 
> > Could it have something to do with whether the machine has a monitor attached? What happens if you reboot your Snow Leopard machine with no monitor attached, ssh in and run the test?
> 
> All the webgl Layout tests crash if I do that.
> 
> I guess I don't follow how not having a monitor attached would affect this test.   Isn't DumpRenderTree pushing everything through Mesa?

only --chromium will run top of mesa.

[Eric Seidel (no email)]

(In reply to comment #15)
> (In reply to comment #14)
> 
> > Could it have something to do with whether the machine has a monitor attached? What happens if you reboot your Snow Leopard machine with no monitor attached, ssh in and run the test?
> 
> All the webgl Layout tests crash if I do that.
> 
> I guess I don't follow how not having a monitor attached would affect this test.   Isn't DumpRenderTree pushing everything through Mesa?

All layout tests will crash when run via SSH. :)  They require an active UI session.  so you have to use VNC.  It may be possible these days to first VNC to the box and then connect via SSH and run them from within the SSH connection while VNC'd.