Bug 69154 – feBlend uses a table of function pointers which reduces inlineability inside the main loop

• Status: RESOLVED
• Resolution: FIXED
• Component: SVG

[Tim Horton]

Before:

1 blend 2083500 pixel image took 8408 us
2 blend 2083500 pixel image took 8576 us
3 blend 2083500 pixel image took 9006 us
4 blend 2083500 pixel image took 9942 us
5 blend 2083500 pixel image took 9592 us

After:

1 blend 2083500 pixel image took 6792 us
2 blend 2083500 pixel image took 6592 us
3 blend 2083500 pixel image took 6896 us
4 blend 2083500 pixel image took 8368 us
5 blend 2083500 pixel image took 7974 us

20% speedup

[Radar WebKit Bug Importer]

<rdar://problem/10215221>

[Tim Horton]

Created attachment 109310 [details]
patch

[Tim Horton]

(In reply to comment #2)
> Created an attachment (id=109310) [details]
> patch + caching bbox

autofill :-\

[WebKit Review Bot]

Attachment 109310 [details] did not pass style-queue:

Failed to run "['Tools/Scripts/check-webkit-style', '--diff-files', u'Source/WebCore/ChangeLog', u'Source/WebCor..." exit_code: 1

Source/WebCore/platform/graphics/filters/FEBlend.cpp:124:  A case label should not be indented, but line up with its switch statement.  [whitespace/indent] [4]
Total errors found: 1 in 2 files


If any of these errors are false positives, please file a bug against check-webkit-style.

[Tim Horton]

Created attachment 109312 [details]
style

[Darin Adler]

Comment on attachment 109312 [details]
style

Putting the switch inside the loop speeds things up by 20%. Could you speed up even more by putting the switch outside the loop? With template functions you could even inline the functions and have the switch statement outside the entire apply function. How much more would that speed things up?

[Zoltan Herczeg]

(In reply to comment #6)
> (From update of attachment 109312 [details])
> Putting the switch inside the loop speeds things up by 20%. Could you speed up even more by putting the switch outside the loop? With template functions you could even inline the functions and have the switch statement outside the entire apply function. How much more would that speed things up?

Probably not much thanks to the advanced branch predictors. But templates probably could be an easy way to do this refactor, and we can see what happens.

[WebKit Review Bot]

Comment on attachment 109312 [details]
style

Clearing flags on attachment: 109312

Committed r96407: <http://trac.webkit.org/changeset/96407>

[WebKit Review Bot]

All reviewed patches have been landed.  Closing bug.

[Tim Horton]

(In reply to comment #7)
> (In reply to comment #6)
> > (From update of attachment 109312 [details] [details])
> > Putting the switch inside the loop speeds things up by 20%. Could you speed up even more by putting the switch outside the loop? With template functions you could even inline the functions and have the switch statement outside the entire apply function. How much more would that speed things up?
> 
> Probably not much thanks to the advanced branch predictors. But templates probably could be an easy way to do this refactor, and we can see what happens.

Seems to be a very small change, if at all; I guess I got it to the point where the compiler could do the rest of the work and hoist the switch out of the loop. I'm testing with release builds; I bet your proposed optimizations would show a bigger change in a debug build.