561385 - Reftest zoomed-1.xhtml is perma-orange on Mac OS 10.6.2 boxes

Status: RESOLVED FIXED

(Core :: Audio/Video, defect)

Description

[Daniel Holbert [:dholbert]]

{
http://tinderbox.mozilla.org/showlog.cgi?log=Firefox/1272038488.1272039436.14387.gz
Rev3 MacOSX Snow Leopard 10.6.2 mozilla-central opt test reftest on 2010/04/23 09:01:28
s: talos-r3-snow-007
REFTEST TEST-UNEXPECTED-FAIL | file:///Users/cltbld/talos-slave/mozilla-central-snowleopard-opt-u-reftest/build/reftest/tests/layout/reftests/ogg-video/zoomed-1.xhtml |
}

Here are some failures that were mistakenly noted on Bug 494420:
http://tinderbox.mozilla.org/showlog.cgi?log=Firefox/1272004384.1272005138.8515.gz
Rev3 MacOSX Snow Leopard 10.6.2 mozilla-central opt test reftest, Started
23:33, finished 23:46
s: talos-r3-snow-003

http://tinderbox.mozilla.org/showlog.cgi?log=Firefox/1272032423.1272033119.28409.gz
Rev3 MacOSX Snow Leopard 10.6.2 mozilla-central opt test reftest, Started
07:20, finished 07:32
s: talos-r3-snow-016

http://tinderbox.mozilla.org/showlog.cgi?log=Firefox/1272035609.1272036472.5513.gz
Rev3 MacOSX Snow Leopard 10.6.2 mozilla-central opt test reftest, Started
08:13, finished 08:28
s: talos-r3-snow-005

Reftest snapshot coming up.

Comment 1

[Daniel Holbert [:dholbert]]

Attachment: reftest failure log

Here's the reftest log.  From glancing in reftest-analyzer.xhtml, the test & reference appear visually the same to me -- both are black rectangles. Maybe they're a slightly different shade of black?

Comment 2

[Daniel Holbert [:dholbert]]

As henri mentioned in bug 561302, this seems to have been introduced by something from this push:
http://hg.mozilla.org/mozilla-central/pushloghtml?changeset=a9ec36359a4c

It's turned that box perma-orange since then.

Comment 4

[Daniel Holbert [:dholbert]]

My guess would be this was caused by Bug 551378 or Bug 551277, since they're about video color conversion.

doublec, any thoughts / obvious bustage fixes?

If I don't hear anything in the next hour or so, I intend to back out both of those bugs to (hopefully) stop the perma-orange.

Comment 5

[Daniel Holbert [:dholbert]]

(In reply to comment #1)
> From glancing in reftest-analyzer.xhtml, the test &
> reference appear visually the same to me -- both are black rectangles. Maybe
> they're a slightly different shade of black?

I just verified this in Gimp -- the testcase is colored rgb(2,2,2), and the reference is rgb(1,1,1).

Comment 6

[Daniel Holbert [:dholbert]]

(In reply to comment #4)
> If I don't hear anything in the next hour or so, I intend to back out both of
> those bugs to (hopefully) stop the perma-orange.

Actually, since they're fairly large pushes (and at least one had a bustage followup or two I think), I just disabled the test on Mac, to stop the random orange:
http://hg.mozilla.org/mozilla-central/rev/3e05788f1ab0

This definitely needs some actual investigation/fixing, though -- I don't intend this test-disabling to last longer than a day or so (maybe a few days given that it's the weekend in New Zealand).

Assigning to cdouble for investigation / backouts / bustage-fixing.

Comment 7

[Robert O'Callahan (:roc) (email my personal email if necessary)]

We probably just need to adjust the reference to be rgb(2,2,2).

Comment 8

[cajbir (:cajbir)]

I think the test only fails on Mac OS X so changing reference will cause failures on the other platforms. Does the Mac test machines use the OpenGL layers YCbCr code or the ycbcr library code?

Comment 9

[Robert O'Callahan (:roc) (email my personal email if necessary)]

The library code. No reftests will be using GL yet.

Maybe we should change the reference to be another video?

Comment 10

[Daniel Holbert [:dholbert]]

(In reply to comment #8)
> I think the test only fails on Mac OS X

More precisely: it fails only on Mac OS X 10.6, I think.

I just updated the reftest.list annotation to reflect this (so we'll know if it starts passing, and we'll also know if it starts failing on other OS X versions)
http://hg.mozilla.org/mozilla-central/rev/61c2144b66df

Comment 11

[cajbir (:cajbir)]

The test passes fine on my Mac OS X 10.6 laptop. I'll try an actually nightly build and see if that behaves any different.

Comment 12

[cajbir (:cajbir)]

I just noticed the bug is marked as x86_64 - does it not happen on 32 bit builds?

Comment 13

[Daniel Holbert [:dholbert]]

I'm not 100% sure, but I think our 10.6 mac boxes are *all* x86_64.  (and our 10.5 ones are all x86)

If that's correct, then maybe it's the 64-bit-ness of those boxes that's causing the problem, and not the OS X version...?

Comment 14

[cajbir (:cajbir)]

Yes, looks like it. 64-bit code on Mac OS X and Windows goes through a different code path. I'm looking at it now.

Comment 15

[cajbir (:cajbir)]

Attachment: fix

The C fallback code is used on 64 bit Mac OS X. The code has a rounding error which causes this test to have pixel values 1 higher than expected.

Comment 16

[Mats Palmgren (inactive)]

Should also remove the "# Fails on 10.6, bug 561385" comment in reftest.list

Comment 17

[cajbir (:cajbir)]

Attachment: Address comment 16

Good catch!

Comment 18

[Chris Pearce [:cpearce (Not reading bugmail)]]

http://hg.mozilla.org/mozilla-central/rev/d349761ddc1e