| Summary: | Contiguous area selection (and fill tool) is TOO SLOW! | ||
|---|---|---|---|
| Product: | [Applications] krita | Reporter: | animtim |
| Component: | Tools | Assignee: | Krita Bugs <krita-bugs-null> |
| Status: | RESOLVED FIXED | ||
| Severity: | normal | CC: | halla, pentalis |
| Priority: | NOR | ||
| Version: | git master (please specify the git hash!) | ||
| Target Milestone: | --- | ||
| Platform: | Unlisted Binaries | ||
| OS: | Linux | ||
| Latest Commit: | Version Fixed In: | ||
| Sentry Crash Report: | |||
| Attachments: | callgrind outpu=t | ||
|
Description
animtim
2011-03-11 12:04:38 UTC
Right... Not sure what we can do here, but let's put these tools on the map for the next performance drive. *** Bug 273859 has been marked as a duplicate of this bug. *** Git commit 5eed87e6aa717ad1ce9ce9093b333bad4680e6c0 by Boudewijn Rempt. Committed on 15/10/2011 at 13:40. Pushed by rempt into branch 'master'. Do not pass a qbitarray of channelflags if they are all true The composite ops optimize when an empty channelflags variable is passed: all channels are then assumed to be turned on. This makes a difference when doing a valgrind run, where before this patch filling an area would spend ~20% of cpu time in testBit and after the patch only ~2%. CCBUG:268212 M +17 -6 krita/image/kis_async_merger.h http://commits.kde.org/calligra/5eed87e6aa717ad1ce9ce9093b333bad4680e6c0 Okay, this is the big, bad boy now:
virtual quint8 difference(const quint8* src1, const quint8* src2) const {
quint8 lab1[8], lab2[8];
cmsCIELab labF1, labF2;
if (this->opacityU8(src1) == OPACITY_TRANSPARENT_U8
|| this->opacityU8(src2) == OPACITY_TRANSPARENT_U8)
return (this->opacityU8(src1) == this->opacityU8(src2) ? 0 : 255);
Q_ASSERT(this->toLabA16Converter());
this->toLabA16Converter()->transform(src1, lab1, 1),
this->toLabA16Converter()->transform(src2, lab2, 1),
cmsLabEncoded2Float(&labF1, (cmsUInt16Number *)lab1);
cmsLabEncoded2Float(&labF2, (cmsUInt16Number *)lab2);
qreal diff = cmsDeltaE(&labF1, &labF2);
if (diff > 255)
return 255;
else
return qint8(diff);
}
And insanely expensive function which takes about 40% of cpu time of doing a fill, according to valgrind. We need a cheaper difference, specialized for the most common colorspaces.
Git commit 551b0e25ed049db65ed36b815447072f544ee38c by Boudewijn Rempt. Committed on 15/10/2011 at 15:14. Pushed by rempt into branch 'master'. Optimize the creation of a flood selection Flood selection uses KoColorSpace::difference which is very, very expensive. Implement a simple cache that works in the most common situation: 8 bit rgba (or any other colorspace where a pixel fits exactly in a quint32). It's still too slow, though. CCBUG: 268212 M +46 -8 krita/image/kis_fill_painter.cc http://commits.kde.org/calligra/551b0e25ed049db65ed36b815447072f544ee38c hm, this callgrind output shows that a) the redisplay for this 4000x4000 images a lot of the time, 40% is spent in the redisplay; the ~30% doing the actual floodfill. The callgrind was generated by started without instrumentation, painting a loop, starting instrumentation, filling, wait until fill had shown up, stopping instrumentation. Created attachment 69179 [details]
callgrind outpu=t
After the ng iterator patch, filling is now pretty fast, I was amazed actually. On this system, there's no difference between gimp 2.8 and krita anymore, though photoshop 7 still is much faster. Animtim, can you check whether the performance is now acceptable for you? I tested with a 3000x3000 image and a loop that enclosed most of the image. Yep, I confirm it's much better already now, good work! (can close this bug I think) great! |