691354 - GL layers on Android: Most reftest failures are caused by pixels values being 'slightly' off

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[Ali Juma [:ajuma]]

Attachment: Always use nearest filtering

With GL layers enabled on Android, several reftests are failing (Bug 676831 flagged many such tests). It turns out that virtually all of these failures are caused by pixel values having one or more channels off by 4 or 8 (e.g., instead of (255,255,255), we get (247,255,247)).

Interestingly, if we require GL to always use nearest (rather than linear) filtering, all but one of the reftest failures are resolved. We wouldn't want to land such a change since, among other things, it makes zooming look ugly, but this might help point us in the direction of the real problem.

Comment 1

[Chris Jones [:cjones] inactive; ni?/f?/r? if you need me]

What happens if you turn off the MOZ_GFX_OPTIMIZE_MOBILE => nearest hacks in bug 688104?

Comment 2

[Ali Juma [:ajuma]]

(In reply to Chris Jones [:cjones] [:warhammer] from comment #1)
> What happens if you turn off the MOZ_GFX_OPTIMIZE_MOBILE => nearest hacks in
> bug 688104?

Even with those hacks turned off, we get the same problem.

Comment 3

[Benoit Girard (:BenWa)]

Attachment: Log for reftest analyzer

Comment 4

[Benoit Girard (:BenWa)]

The problem seems to be caused by the use of LINEAR filter with TiledTextureImage with large textures. The following patch changes the leakage to a clear pattern on the right edge of the rectangle.

-    mTileSize = mGL->GetMaxTextureSize();
+    mTileSize = 512;//mGL->GetMaxTextureSize();

Comment 5

[Benoit Girard (:BenWa)]

Attachment: Log with 512 max TiledTextureImage

Comment 6

[Benoit Girard (:BenWa)]

And the test passes with a 256 tile size.

Comment 7

[Benoit Girard (:BenWa)]

Attachment: ShaderHacking

Posting some shader hacking that I worked on with Jeff to get an idea of what is going on. Our general belief is that it is related to precision but we don't understand the problem.

Comment 8

[Benoit Girard (:BenWa)]

Attachment: patch

More of a work around then a fix unfortunately.

Comment 9

[Benoit Girard (:BenWa)]

Attachment: Enable test fix by patch

With these two patch applied we fail 3 tests and pass 3 test that !layerOpenGL fails.

~/mozilla/mozilla-central/tree> find . -name "reftest.list" | xargs grep "Android&&"
./layout/reftests/text-overflow/reftest.list:fails-if(Android&&layersOpenGL) HTTP(..) == block-padding.html block-padding-ref.html
./layout/reftests/canvas/reftest.list:fails-if(Android&&layersOpenGL) == image-rendering-test.html image-rendering-ref.html
./layout/reftests/canvas/reftest.list:fails-if(Android&&layersOpenGL) == image-shadow.html image-shadow-ref.html
./layout/reftests/svg/reftest.list:fails-if(Android&&!layersOpenGL) == dynamic-conditions-01.svg pass.svg # bug 652050
./layout/reftests/svg/reftest.list:fails-if(Android&&!layersOpenGL) == dynamic-switch-01.svg pass.svg # bug 652050
./layout/reftests/svg/reftest.list:fails-if(Android&&!layersOpenGL) == switch-01.svg pass.svg # bug 652050

Comment 10

[Benoit Girard (:BenWa)]

Attachment: patch v2 (handle 3d transforms)

Comment 11

[Benoit Girard (:BenWa)]

Attachment: patch v3 (handle 3d transforms)

Comment 12

[Benoit Girard (:BenWa)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/5061db810f1f

Comment 13

[Marco Bonardo [:mak] (Away Apr 25 - May 5)]

https://hg.mozilla.org/mozilla-central/rev/5061db810f1f
https://hg.mozilla.org/mozilla-central/rev/cd9fe4275983

Comment 14

[Benoit Girard (:BenWa)]

Still need to land test changes. I was waiting on TBPL results to compare my local results with tinderbox, currently the tree is closed.

Comment 15

[Benoit Girard (:BenWa)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/a456c129a2c3

Comment 16

[Marco Bonardo [:mak] (Away Apr 25 - May 5)]

https://hg.mozilla.org/mozilla-central/rev/a456c129a2c3

Comment 17

[Marco Bonardo [:mak] (Away Apr 25 - May 5)]

merged follow-up as well https://hg.mozilla.org/mozilla-central/rev/6688e7c9cd8f