Bug 78323 - [chromium] Increase JPEG decoding performance some moar
Summary: [chromium] Increase JPEG decoding performance some moar
Status: RESOLVED FIXED
Alias: None
Product: WebKit
Classification: Unclassified
Component: New Bugs (show other bugs)
Version: 528+ (Nightly build)
Hardware: Unspecified Unspecified
: P2 Normal
Assignee: noel gordon
URL:
Keywords:
Depends on: 59670 75861 78454 85772
Blocks:
  Show dependency tree / graph
 
Reported: 2012-02-09 20:47 PST by noel gordon
Modified: 2012-05-11 03:32 PDT (History)
6 users (show)

See Also:

Attachments
Patch (19.72 KB, patch)
2012-02-09 21:14 PST, noel gordon 		no flags Details | Formatted Diff | Diff
tjbench.fast.r778.svn.diff (1.66 KB, text/plain)
2012-02-09 22:23 PST, noel gordon 		no flags Details
View All Add an attachment (proposed patch, testcase, etc.)

Note You need to log in before you can comment on or make changes to this bug.
Description noel gordon 2012-02-09 20:47:09 PST 
[chromium] Increase JPEG decoding performance some moar
Comment 1 noel gordon 2012-02-09 21:14:05 PST 
Created attachment 126447 [details]
Patch
Comment 2 noel gordon 2012-02-09 22:21:46 PST 
r101286 increased JPEG decoding performance by ~2x for the chromium ports, and similarly for the webkit android port in r104472.  I then looked for even more performance and the result is this patch.

Created an corpus of 369 jpeg images, composed of 4:4:4, 4:2:2, 4:2:0 chroma sampled images in graysacle, cmyk, and rgb (the majority). Modified libjpeg-turbo tjbench.c to allow the selection of fast decoding via a command line option. Compiled tjbench under cygwin windows7 x86_64 on a T410 laptop and a Z600 desktop, and also on a MacPro x86_32.  Modified tjbench svn.diff at libjpeb-turbo r778 attached.
Comment 3 noel gordon 2012-02-09 22:23:33 PST 
Created attachment 126454 [details]
tjbench.fast.r778.svn.diff
Comment 4 noel gordon 2012-02-09 22:25:03 PST 
Ran "./tjbench image.jpg -bgrx" for each image in the corpus.  tjbench continuously decodes the image for 5 seconds, then reports the decoding throughput (Mbyte/s).  I then repeated the experiment with fast decoding.  The results are tabulated and graphed in the following spreadsheet.

  https://docs.google.com/spreadsheet/ccc?key=0Asd5UUPIMYYsdFNyTW95YzFUazJtUUZYY25UM0lZQ2c

The results suggest fast decoding provides a 9% speed improvement (avg) across all machines.
Comment 5 Adam Barth 2012-02-09 23:39:02 PST 
Comment on attachment 126447 [details]
Patch

Looks reasonable assuming you've reviewed all the changed test results as reasonable.
Comment 6 noel gordon 2012-02-10 00:02:00 PST 
I have reviewed, and I am satisfied with, the changed test results on win and mac chrome.
Comment 7 WebKit Review Bot 2012-02-10 02:46:27 PST 
Comment on attachment 126447 [details]
Patch

Clearing flags on attachment: 126447

Committed r107389: <http://trac.webkit.org/changeset/107389>
Comment 8 WebKit Review Bot 2012-02-10 02:46:32 PST 
All reviewed patches have been landed.  Closing bug.
Comment 9 Pavel Podivilov 2012-02-10 09:15:19 PST 
Broke ExtensionChromePrivateApiTest.DecodeJPEG downstream.

http://test-results.appspot.com/dashboards/flakiness_dashboard.html#testType=browser_tests&tests=ExtensionChromePrivateApiTest.DecodeJPEG
Comment 10 noel gordon 2012-02-10 14:25:47 PST 
ExtensionChromePrivateApiTest.DecodeJPEG compares image bytes, rebaseline it after the webkit roll.
Comment 11 noel gordon 2012-02-10 14:31:15 PST 
Peter, Sami: don't know how you rebaseline android layout test results, but you'll need to when this change rolls into chrome android.
Comment 12 noel gordon 2012-02-19 23:01:50 PST 
Desktop chrome rebaselines, bug 78454.
Comment 13 noel gordon 2012-05-11 02:13:55 PDT 
Backed part of this out in bug 85772.  We now only allow IFAST decodes on small screen devices (Chromium Android).