Bug 19312

Summary: SVG hit testing is *way* too slow
Product: WebKit Reporter: Eric Seidel (no email) <eric>
Component: SVGAssignee: Nobody <webkit-unassigned>
Status: RESOLVED FIXED    
Severity: Normal CC: alp, krit, simon.fraser, zimmermann
Priority: P2    
Version: 528+ (Nightly build)   
Hardware: Mac   
OS: OS X 10.5   
Attachments:
Description Flags
test case (sorta a layout test, not quite)
none
Patch
none
Patch none

Eric Seidel (no email)
Reported 2008-05-29 10:57:43 PDT
SVG hit testing is *way* too slow See the attached test case. I think the major problem is that RenderPath::nodeAtPoint() does not to any sort of optimization (like checking against the bbox) before testing against fill contains. Thus for an example like the one in the test case, we end up doing hundreds if not thousands of unnecessary path containment tests (which may involve mallocs, etc).
Attachments
test case (sorta a layout test, not quite) (2.54 KB, image/svg+xml)
2008-05-29 10:58 PDT, Eric Seidel (no email) 		no flags
Details
Patch (1.57 KB, patch)
2008-12-30 22:32 PST, Simon Fraser (smfr) 		no flags
DetailsFormatted DiffDiff
Patch (1.59 KB, patch)
2010-05-04 12:49 PDT, Dirk Schulze 		no flags
DetailsFormatted DiffDiff
Show Obsolete (1) View All Add attachment proposed patch, testcase, etc.
Eric Seidel (no email)
Comment 1 2008-05-29 10:58:59 PDT
Created attachment 21416 [details] test case (sorta a layout test, not quite) See the attached test case. We don't really have a very good way to do perf-based layout tests. The attached test could be modified to be say PASS when it runs under a certain threshold or something.
Simon Fraser (smfr)
Comment 2 2008-12-30 22:32:27 PST
Created attachment 26329 [details] Patch This patch implements your suggestion, and removes much of the hit testing cost. However, repaint is still very expensive. Sharking found bug 23050.
Simon Fraser (smfr)
Comment 3 2008-12-30 22:44:47 PST
Repaint is very expensive because EVERYTHING repaints even if only a small area is dirtied.RenderPath::paint() does no short-circuiting if the relativeBBox() is not in the dirty rect.
Simon Fraser (smfr)
Comment 4 2009-01-02 15:30:08 PST
See also bug 15393.
Dirk Schulze
Comment 5 2009-12-30 13:36:02 PST
*** Bug 15704 has been marked as a duplicate of this bug. ***
Dirk Schulze
Comment 6 2009-12-30 13:44:11 PST
(In reply to comment #3) > Repaint is very expensive because EVERYTHING repaints even if only a small area > is dirtied.RenderPath::paint() does no short-circuiting if the relativeBBox() > is not in the dirty rect. Will the patch on bug 30055 fix this?
Dirk Schulze
Comment 7 2010-05-04 12:49:40 PDT
Created attachment 55039 [details] Patch
Dirk Schulze
Comment 8 2010-05-04 12:50:58 PDT
(In reply to comment #7) > Created an attachment (id=55039) [details] > Patch This is a huge speed-up of the test attached to this br.
Nikolas Zimmermann
Comment 9 2010-05-04 13:06:48 PDT
Comment on attachment 55039 [details] Patch Looks great to me, as discussed before :-) r=me.
Dirk Schulze
Comment 10 2010-05-05 00:31:46 PDT
Comment on attachment 55039 [details] Patch Clearing flags on attachment: 55039 Committed r58804: <http://trac.webkit.org/changeset/58804>
Dirk Schulze
Comment 11 2010-05-05 00:31:57 PDT
All reviewed patches have been landed. Closing bug.
Note You need to log in before you can comment on or make changes to this bug.