[Insight-users] [offlist] SignedMaurerDistanceMapImageFilter performance

[Sergio Vera]

Hi Bradley!

Indeed the snippet of code used nonsense images (I tried to remove the
conversion from unsigned char to float from the equation).
I've tested ITK 4.1 with a small improvement from 13 to 11 seconds.
But your pacth improved the results to merely 6 seconds of time, bringing
the times 50% down.

Many thanks

On Tue, Jul 3, 2012 at 5:31 AM, Bradley Lowekamp <blowekamp at mail.nih.gov>wrote:

> Hello Sergio,
>
> I did also notice that you have a different output pixel type than I.
> Simple ITK was using float. As you are using output pixel type uchar, but
> with using image spacing, I am not sure how much sense that combination
> makes. Additionally you have signed output into uchar. This may be part for
> the problem. I tried changing the output pixel type in simpleITK to the
> same as the input, and I got similar to slightly faster results.
>
> Looking at the code, I did see a multi-threadhing issue, I have made a
> patch for it, that is still a work in progress:
>
> http://review.source.kitware.com/#/c/6367/
>
> I haven't tried running your code yet. The performance difference still
> seems very off to me.
>
> Brad
>
>
> On Jul 2, 2012, at 12:15 PM, Sergio Vera wrote:
>
> Hi  Dženan and Bradley,
>
> I use Control+F5 normally. and it's a Professional version of Visual
> Studio 9 (2008)
>
> I've made a minimum test scenario to remove any possible interferences
> from other libraries/dependencies. This one only links against ITK, and
> still times are arrount 13-14 segs.
> By the way switching to ITK 4.0 did not improve my numbers :(
>
> cmake_minimum_required(VERSION 2.6)
> project(dmap)
> find_package(ITK REQUIRED)
> include(${ITK_USE_FILE})
> add_executable(dmap main.cxx)
> target_link_libraries(dmap  ${ITK_LIBRARIES})
>
> #include <itkTimeProbe.h>
> #include <itkImage.h>
> #include <itkImageFileReader.h>
> #include <itkSignedMaurerDistanceMapImageFilter.h>
>
> int main(int argc, char *argv[])
> {
> typedef itk::Image<unsigned char,3> ITKImg;
> typedef itk::ImageFileReader<ITKImg> Reader;
>  Reader::Pointer reader = Reader::New();
>
> reader->SetFileName("C:/Users/sergio.vera/Documents/MATLAB/images/CTAbdomenThr1800.mha");
>  reader->Update();
> std::cout << "Image readed!\n ";
> itk::TimeProbe timer;
>  timer.Start();
> typedef itk::SignedMaurerDistanceMapImageFilter <ITKImg, ITKImg>
> DistanceMapper;
>  DistanceMapper::Pointer distanceMapper = DistanceMapper::New();
> distanceMapper->SetUseImageSpacing(true);
>  distanceMapper->SetSquaredDistance(false);
> distanceMapper->SetInsideIsPositive(false);
>  distanceMapper->SetInput(reader->GetOutput());
> distanceMapper->SetNumberOfThreads(2);
>  distanceMapper->Update();
>  timer.Stop();
>  std::cout << "Time Spent: " << timer.GetMeanTime() << "\n";
>   return EXIT_SUCCESS;
> }
>
>
> On Mon, Jul 2, 2012 at 6:02 PM, Bradley Lowekamp <blowekamp at mail.nih.gov>wrote:
>
>> Would you by any chance be using an express Visual Studio version? Those
>> compiler are missing a fair amount of optimization. I have not compared the
>> difference myself though.
>>
>> Brad
>>
>> On Jul 2, 2012, at 11:38 AM, Sergio Vera wrote:
>>
>> Hello Bradley
>>
>> Thanks for your script
>>
>> I've tried your script and the timings are coherent with yours (although
>> my desktop is slower than your laptop :) )
>>
>> The average time it took to execute SignedMaurerDistanceMap was
>> 5.3353600223 seconds.
>> The average time it took to execute SignedMaurerDistanceMap was
>> 6.62535580948 seconds.
>> The average time it took to execute SignedMaurerDistanceMap was
>> 6.90300935124 seconds.
>>
>> I've modified it to read my image (it's only 300k but it seems too big
>> for the list) and the results are definitely lower than 13 seconds, closer
>> to the 7 seconds from Mevislab's ITK....
>>
>> The average time it took to execute SignedMaurerDistanceMap was
>> 8.87152283895 seconds.
>> The average time it took to execute SignedMaurerDistanceMap was
>> 9.38857726434 seconds.
>> The average time it took to execute SignedMaurerDistanceMap was
>> 9.54250972999 seconds.
>>
>> Indeed it is strange that my code runs slower :( My executable being
>> compiled in release mode, compiler flags are /02 and Visual studio is
>> clearly put in release. Also, when I launch the debug version the distance
>> map time is 114 seconds. However I'm not sure why it tools so long to
>> compute the distance map..
>>
>> Thanks
>>
>> On Mon, Jul 2, 2012 at 4:07 PM, Bradley Lowekamp <blowekamp at mail.nih.gov>wrote:
>>
>>> Hello Sergio,
>>>
>>> But are you compiling your program also in release mode? The filter is
>>> not part of the ITK code in the library, because it is a templated filter,
>>> it will be compiled in your application, so that is where it matters if you
>>> are in Release or debug mode.
>>>
>>> I wrote the following little SimpleITK python script, which uses the
>>> latest ITK 4.2rc04:
>>>
>>> import SimpleITK as sitk
>>> import os
>>> from timeit import Timer
>>>
>>> repeat = 10
>>>
>>> img = sitk.Image( [512,512,110], sitk.sitkUInt8 )
>>> img[255,255,55 ] = 1
>>>
>>> filter = sitk.SignedMaurerDistanceMapImageFilter()
>>>
>>> print "Defaults:"
>>> print filter
>>>
>>> t = Timer( lambda: filter.Execute( img ) )
>>>
>>> print "The average time it took to execute", filter.GetName(), "was",
>>>  min(t.repeat( repeat=repeat,number=1 )), "seconds. "
>>>
>>> filter.InsideIsPositiveOff()
>>> filter.SquaredDistanceOff()
>>> filter.UseImageSpacingOn();
>>>
>>> t = Timer( lambda: filter.Execute( img ) )
>>>
>>> print "The average time it took to execute", filter.GetName(), "was",
>>>  min(t.repeat( repeat=repeat,number=1 )), "seconds. "
>>>
>>> t = Timer( lambda: filter.Execute( ~img ) )
>>>
>>> print "The average time it took to execute", filter.GetName(), "was",
>>>  min(t.repeat( repeat=repeat,number=1 )), "seconds. "
>>>
>>> Your large image didn't come attached, so I just created a trivial on to
>>> run performance. This is my results in my i7 laptop:
>>>
>>> Defaults:
>>> itk::simple::SignedMaurerDistanceMapImageFilter
>>>   InsideIsPositive: 0
>>>   SquaredDistance: 1
>>>   UseImageSpacing: 0
>>>
>>> The average time it took to execute SignedMaurerDistanceMap was
>>> 2.05940294266 seconds.
>>> The average time it took to execute SignedMaurerDistanceMap was
>>> 2.61217999458 seconds.
>>> The average time it took to execute SignedMaurerDistanceMap was
>>> 2.7291328907 seconds.
>>>
>>>
>>> As I was running this multiple times, I appear to be getting a variation
>>> of ~.5 seconds due to what I am assuming is the temperature of my CPU, and
>>> intel's TuboBoost.
>>>
>>> Additionally, it's not clear which timings are using only one thread and
>>> which are using more. So I also ran the test with only one thread:
>>>
>>> The average time it took to execute SignedMaurerDistanceMap was
>>> 8.77108097076 seconds.
>>> The average time it took to execute SignedMaurerDistanceMap was
>>> 9.40146708488 seconds.
>>> The average time it took to execute SignedMaurerDistanceMap was
>>> 9.63690400124 seconds.
>>>
>>> I tend to agree with Bill, there is something funny about how you
>>> compiled it.
>>>
>>> Brad
>>>
>>> _____________________________________
>>> Powered by www.kitware.com
>>>
>>> Visit other Kitware open-source projects at
>>> http://www.kitware.com/opensource/opensource.html
>>>
>>> Kitware offers ITK Training Courses, for more information visit:
>>> http://www.kitware.com/products/protraining.php
>>>
>>> Please keep messages on-topic and check the ITK FAQ at:
>>> http://www.itk.org/Wiki/ITK_FAQ
>>>
>>> Follow this link to subscribe/unsubscribe:
>>> http://www.itk.org/mailman/listinfo/insight-users
>>>
>>
>>
>>
>> --
>> Sergio Vera
>>
>>  Alma IT Systems
>>  C/ Vilana, 4B, 4º 1ª
>>  08022 Barcelona
>>  T. (+34) 932 380 592
>>  www.alma3d.com
>>
>>
>>  ========================================================
>> Bradley Lowekamp
>> Medical Science and Computing for
>> Office of High Performance Computing and Communications
>> National Library of Medicine
>>  blowekamp at mail.nih.gov
>>
>>
>>
>>
>
>
> --
> Sergio Vera
>
>  Alma IT Systems
>  C/ Vilana, 4B, 4º 1ª
>  08022 Barcelona
>  T. (+34) 932 380 592
>  www.alma3d.com
>
>
>


-- 
Sergio Vera

 Alma IT Systems
 C/ Vilana, 4B, 4º 1ª
 08022 Barcelona
 T. (+34) 932 380 592
 www.alma3d.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-users/attachments/20120703/a13b997b/attachment.htm>