Dear Sacha, just synced to the current master with the factoryless new macro branch merged. The performance just skyrocketed for me both in debug and release versions! Thanks a lot!
Rostislav. On 13/02/2014 17:08, Sascha Zelzer wrote: > Hi, > > thanks a lot for sharing your experience and evaluation. > > Your are right that the standard ::New() call adds a lot of overhead > and the provided flexibility (asking object factories) is rarely > needed. We indeed discussed the possibility to replace most (if not > all) ::New() methods in MITK with the "factory-less" version. We > actually agreed in general that this should be done but no action has > been carried out so far. > > You are also right that bug 14866 will drastically reduce the amount > of registered object factories and should result in an overall New() > speed-up. > > Concerning the lazy Geometry2D evaluation in SlicedGeometry3D and the > Clone() overhead, your ideas make totally sense. A student of ours is > currently investigating the class hierarchy of the geometry classes > also the Geometry2D / SlicedGeometry3D relationship with the goal of a > simplified hierarchy. We will certainly take your experiences into > account! > > So work on the factories and geometries is already being carried out > and the results should improve your runtime performance (timeframe: > ~3months). We will also pick up the factory-less New() discussion again. > > Thanks, > Sascha > > On 02/12/2014 05:21 PM, Rostislav Khlebnikov wrote: >> Hello, >> >> I would like to share some of my experiences working with geometries in >> MITK and some more or less general concerns that came out of them. >> >> Right now I am implementing a rather common thing - reslicing the image >> perpendicularly to a curve (vessel path). >> I have found a very old discussion about the similar task: >> http://sourceforge.net/mailarchive/message.php?msg_id=3457811 >> Even though it was from 2007, it seemed quite applicable to the current >> MITK version as well with some minor modifications. >> >> So the first approach was to create the necessary PlaneGeometry >> instances, feed it to a SlicedGeometry3D and set this geometry to the >> renderer's SliceNavigationController. >> It does indeed work as intended. However, the creation of slices is >> terribly slow: 4-5 seconds for 50 slices! Given that the slices might be >> required once every user click (e.g. the vessel path changes) this makes >> the application unusable. >> I have experimented a little trying to find the reason why is it this >> slow and this is what I found. >> 1) <DataType>::New() is relatively slow because the call goes through >> the itk factories checking for overrides. In my current version of MITK >> Workbench with quite some plugins enabled - it is 81 factories which are >> checked. >> 2) All the geometries are cloned at least once upon setting the geometry >> to the SNC (which also works through <DataType>::New() and the time >> spent in allocating the objects is at least doubled). >> >> I have made a small test to check the timings of creating a number of >> plane geometries. This is what I found: >> 1) Creating a plane geometry with all mitk factories registered takes >> about 5.5ms (*) >> 2) Creating a plane geometry without any mitk factories registered (only >> builtin ITK factories registered - ~11 of them) takes about 1.1ms >> 3) Substituting the itkNewMacro() with itkFactorylessNewMacro() in all >> the geometry classes makes the creation time to be about 0.9ms >> These tests are obviously not conclusive and more extensive and >> controlled testing would be needed, but I think it shows the overall >> picture. >> >> (*) Note, that the timings were taken in debug mode. However, when I >> compiled the first (non-test) version of my code in release the creation >> of 50 slices still took about 3 seconds, so I assume that in this >> particular case Release optimizations do not affect performance >> significantly. >> >> Seems to me that at least one of the reasons for SlicedGeometry3D having >> special cases for lazy generation of Geometry2Ds in case it is evenly >> spaced is low performance of New(). >> Anyway, I wanted to draw the attention of MITK core developers to this >> problem as I am concerned that it may slow down the overall MITK >> performance unnecessarily. >> >> I am unsure whether this work will reduce the number of registered mitk >> factories - http://bugs.mitk.org/show_bug.cgi?id=14866. >> >> While I think that improving the New() performance is important overall, >> I feel like creating all the plane geometries at once is still not the >> best approach to my particular problem (reslicing perpendicular to the >> curve). >> For now I decided to introduce lazy evaluation of Geometry2D's for my >> case by subclassing the SlicedGeometry3D. It works more or less fine, >> but I had to do some hacky things in copy constructor (because the >> SlicedGeometry3D has assertions that check if all Geometry2D's are not >> null if it is not evenly spaced, which is obviously not true for my >> case). However, I cannot avoid subclassing SlicedGeometry3D as it is >> used (through dynamic_cast) in the BaseRenderer. >> >> I would propose to remove the specific implementation of lazy evaluation >> of Geometry2Ds for evenly spaced case from SlicedGeometry3D. It may be >> done by adding another hierarchy level (e.g. having >> EvenlySlicedGeometry3D : public SlicedGeometry3D which overrides >> GetGeometry2D) or by having something like a Geometry2DGenerator member >> which can then be subclassed for each type of lazy evaluation approach >> (e.g. EvenlySpacedSlicesGenerator which is able to create the slices >> using a reference slice, FullyStoredSlicesGenerator which can store all >> the possible slices, or in my case CurvePerpendicularSlicesGenerator). >> This is of course up to the core developers to choose the approach as I >> might miss some important use-cases for the SlicedGeometry3D. >> >> I am really looking forward to hearing your thoughs on these topics. >> If I missed some good solution for reslicing that is already possible in >> MITK - please give me a hint about it. >> >> Rostislav. >> >> PS: >> I know this is quite some text here, but I tried to give as much >> (useful?) information as I could :) >> >> >> ------------------------------------------------------------------------------ >> >> >> Android apps run on BlackBerry 10 >> Introducing the new BlackBerry 10.2.1 Runtime for Android apps. >> Now with support for Jelly Bean, Bluetooth, Mapview and more. >> Get your Android app in front of a whole new audience. Start now. >> http://pubads.g.doubleclick.net/gampad/clk?id=124407151&iu=/4140/ostg.clktrk >> >> _______________________________________________ >> mitk-users mailing list >> [email protected] >> https://lists.sourceforge.net/lists/listinfo/mitk-users > ------------------------------------------------------------------------------ Flow-based real-time traffic analytics software. Cisco certified tool. Monitor traffic, SLAs, QoS, Medianet, WAAS etc. with NetFlow Analyzer Customize your own dashboards, set traffic alerts and generate reports. Network behavioral analysis & security monitoring. All-in-one tool. http://pubads.g.doubleclick.net/gampad/clk?id=126839071&iu=/4140/ostg.clktrk _______________________________________________ mitk-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/mitk-users
