Bug 98647

Summary: Composited/HW content have the red and blue channels inverted in DRT on Chromium Android
Product: WebKit Reporter: Peter Beverloo <peter>
Component: Tools / TestsAssignee: Sami Kyöstilä <skyostil>
Status: RESOLVED FIXED    
Severity: Normal CC: alokp, bsalomon, jamesr, senorblanco, skyostil, tomhudson, webkit.review.bot
Priority: P2    
Version: 528+ (Nightly build)   
Hardware: Unspecified   
OS: Unspecified   
Bug Depends on:    
Bug Blocks: 96398, 100263    
Attachments:
Description Flags
Patch none

Description Peter Beverloo 2012-10-08 06:45:15 PDT 
Composited/HW content have the red and blue channels inverted in DRT on Chromium Android. To list a few examples:

* http://build.webkit.org/results/Chromium%20Android%20Release%20(Tests)/r130632%20(1228)/css3/filters/effect-drop-shadow-hw-diffs.html
    Test: http://trac.webkit.org/export/130632/trunk/LayoutTests/css3/filters/effect-drop-shadow-hw.html

* http://build.webkit.org/results/Chromium%20Android%20Release%20(Tests)/r130632%20(1228)/compositing/shadows/shadow-drawing-diffs.html
    Test: http://trac.webkit.org/export/130632/trunk/LayoutTests/compositing/shadows/shadow-drawing.html

This also seems to happen to text content, such as in this result:

* http://build.webkit.org/results/Chromium%20Android%20Release%20(Tests)/r130632%20(1228)/compositing/sibling-positioning-diffs.html
    Test: http://trac.webkit.org/export/130632/trunk/LayoutTests/compositing/sibling-positioning.html

Normal images and texts are fine. As a point of interest here, the red and blue channels are being inverted by code in DumpRenderTree itself, but it seems that normal (non-accelerated) screenshots also take this path.

http://code.google.com/searchframe#OAMlx_jo-ck/src/third_party/WebKit/Tools/DumpRenderTree/chromium/TestShell.cpp&l=676

Could it be that the SkCanvas in m_webViewHost->canvas() returns BGRA for hardware, RGBA for software paths?
Comment 1 Brian Salomon 2012-10-08 07:13:30 PDT 
I don't think it is a Skia issue as the SW skia backend is used for at least the third test (and possibly the second as well). I'd guess it has to do with invoking the compositor.
Comment 2 Stephen White 2012-10-08 10:41:07 PDT 
I seem to remember Alok ran into a channel swap in the compositor someplace.  Don't know if it was related.
Comment 3 Sami Kyöstilä 2012-10-09 08:40:13 PDT 
I'm guessing there's a mismatch between the compositor's read back pixel format and what Skia/WebViewHost is expecting. We've had a few of these before so it's not entirely unlikely.
Comment 4 Sami Kyöstilä 2012-10-25 08:10:12 PDT 
Created attachment 170652 [details]
Patch
Comment 5 James Robinson 2012-10-25 12:21:26 PDT 
Comment on attachment 170652 [details]
Patch

Hmm, so the byte order for kARGB_8888 is different on android, or the configuration itself is different?
Comment 6 Brian Salomon 2012-10-25 12:40:28 PDT 
The macros SK_R32_SHIFT, SK_G32_SHIFT, etc that define the byte order of Skia's kARGB_8888 are set differently on Android than on other platforms.
Comment 7 James Robinson 2012-10-25 12:47:22 PDT 
Comment on attachment 170652 [details]
Patch

OK, thanks for the clarification.  Patch seems fine but I have a sneaking suspicious that this might not be the only place we have byte order problems.
Comment 8 WebKit Review Bot 2012-10-25 13:01:18 PDT 
Comment on attachment 170652 [details]
Patch

Clearing flags on attachment: 170652

Committed r132518: <http://trac.webkit.org/changeset/132518>
Comment 9 WebKit Review Bot 2012-10-25 13:01:22 PDT 
All reviewed patches have been landed.  Closing bug.