848936 - Various SVG ref test failures on Windows 8

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[Jim Mathies [:jimm]]

The data images don't look different via a visual inspection. 

Guessing this is graphics/d2d related. 

standalone--auto-auto--pct-pct.svg | max difference: 1, differing pixels: 392
standalone--auto-pct--pct-0.svg | max difference: 1, differing pixels: 392
standalone--auto-pct--pct-pct.svg | max difference: 1, differing pixels: 392
standalone--auto-pct--pct-px.svg | max difference: 1, differing pixels: 392
standalone--auto-pct--px-px.svg | max difference: 1, differing pixels: 979
standalone--pct-auto--0-pct.svg | max difference: 1, differing pixels: 392
standalone--pct-auto--pct-pct.svg | max difference: 1, differing pixels: 392
standalone--pct-auto--px-pct.svg | max difference: 1, differing pixels: 392
standalone--pct-pct--0-0.svg | max difference: 1, differing pixels: 392
standalone--pct-pct--0-pct.svg | max difference: 1, differing pixels: 392
standalone--pct-pct--0-px.svg | max difference: 1, differing pixels: 392
standalone--pct-pct--pct-0.svg | max difference: 1, differing pixels: 392
standalone--pct-pct--pct-pct.svg | max difference: 1, differing pixels: 392
standalone--pct-pct--pct-px.svg | max difference: 1, differing pixels: 392
standalone--pct-pct--px-0.svg | max difference: 1, differing pixels: 392
standalone--pct-pct--px-pct.svg | max difference: 1, differing pixels: 392
standalone--pct-pct--px-px.svg | max difference: 1, differing pixels: 392
feColorMatrix-1.svg | max difference: 1, differing pixels: 6400
feColorMatrix-2.svg | max difference: 1, differing pixels: 10000
feComposite-1.svg | max difference: 1, differing pixels: 9600
feComposite-2.svg | max difference: 1, differing pixels: 10000
feFlood-1.svg | max difference: 1, differing pixels: 6400
feFlood-2.svg | max difference: 1, differing pixels: 250000
clipPath-and-shape-rendering-01.svg | max difference: 1, differing pixels: 3
dynamic-filter-contents-01a.svg | max difference: 1, differing pixels: 10000
dynamic-filter-contents-01b.svg | max difference: 1, differing pixels: 10000
mask-type-01.svg | max difference: 1, differing pixels: 6400
mask-type-02.svg | max difference: 1, differing pixels: 6400
mask-type-03.svg | max difference: 1, differing pixels: 6400
mask-type-04.svg | max difference: 1, differing pixels: 6400
opacity-and-gradient-02.svg | max difference: 1, differing pixels: 200000

Comment 1

[Jim Mathies [:jimm]]

Looking specifically at opacity-and-gradient-02.svg, the middle color bar has rgb values of 128,128,0. In the live svg image the resulting image surface we compare has a middle band that transitions half way through from 128,128,0 to 128,127,0. Hence the failure.

Comment 2

[Jim Mathies [:jimm]]

Attachment: color bands

Comment 3

[Jim Mathies [:jimm]]

cc'ing a couple gfx folks. Joe, Jeff, any suggestions on the best place in the code to look for the cause of something like this? Our svg ref tests are failing pretty hard on Win8. The cause appears to be minor color shifting in the initial canvas rendering or in the resulting image surface we do the comparison with.

Comment 4

[Jim Mathies [:jimm]]

Setting layers.acceleration.disabled to true fixes the failure.

Comment 5

[Jim Mathies [:jimm]]

The drivers we have on these are one rev out out of date. I'm not sure if a bad driver could cause something like this though.


          Card name: NVIDIA GeForce GT 430
       Manufacturer: NVIDIA
          Chip type: GeForce GT 430
           DAC type: Integrated RAMDAC
        Device Type: Full Device
         Device Key: Enum\PCI\VEN_10DE&DEV_0DE1&SUBSYS_14303842&REV_A1
     Display Memory: 4050 MB
   Dedicated Memory: 978 MB
      Shared Memory: 3071 MB
       Current Mode: 1600 x 1200 (32 bit) (60Hz)
        Driver Name: nvd3dumx.dll,nvwgf2umx.dll,nvwgf2umx.dll,nvd3dum,nvwgf2um,nvwgf2um
Driver File Version: 9.18.0013.1090 (English)
     Driver Version: 9.18.13.1090
        DDI Version: 11
     Feature Levels: 11.0,10.1,10.0,9.3,9.2,9.1
       Driver Model: WDDM 1.2
Graphics Preemption: DMA
 Compute Preemption: DMA
  Driver Attributes: Final Retail
   Driver Date/Size: 12/29/2012 02:34:47, 18054312 bytes

Comment 6

[Jim Mathies [:jimm]]

Bas, any ideas here as to what might be causing the color shifting in these tests? We could try upgrading the driver if you think that might help. I'm curious if we've run into this kind of problem before with direct2d.

Comment 7

[Bas Schouten (:bas.schouten)]

We discussed this on IRC.

Comment 8

[Jim Mathies [:jimm]]

<jimm> Bas: do you have any ideas on bug 848936? that's holding up getting win8 desktop running on mc.
<Bas> jimm: Those fail for me locally on m-c too.
<Bas> jimm: Our tinderbox test machines are the magical machines where reftests pass, I've never seen them pass with D2D on any other hardware :)
<jimm> Bas: what do you think we should do? fuzzy the heck out of the tests, disable them, ..?
<Bas> jimm: Fuzzy them, yes. A subpixel difference of 1 max is really always innocent when it comes to how we do reftests.
<Bas> jimm: fwiw, I think they're due to GPU rounding not being IEEE 754 compliant

I will work up a patch here and in bug 850457.

Comment 9

[Jim Mathies [:jimm]]

Attachment: patch

annotation for various tests for the new testing hardware. on the few that failed hard, I just skipped.

Comment 10

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

Comment on attachment 724643 [details] [diff] [review]
patch

Review of attachment 724643 [details] [diff] [review]:
-----------------------------------------------------------------

Please file bug(s) for the tests that fail completely.

Comment 11

[Jim Mathies [:jimm]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/071f20163360

Comment 12

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/071f20163360