[Insight-developers] Template Meta Programming and itkImageBase

[Bill Lorensen]

Performance tests I did a while ago:
http://www.itk.org/Wiki/Proposals:Orientation#Recursive_Templates_for_fast_computation_of_Index-Point_transform
 show a large difference between the use of recursive templates and
compiler optimized code.

I'll try top resurrect these tests.

Bill

On Sun, Oct 19, 2008 at 10:03 AM, Luis Ibanez <luis.ibanez at kitware.com> wrote:
>
> Bill,
>
>
>      Agree,
>      this was mostly a documentation issue.
>
>
> The previous implementation of the code had the expression:
>
>
>   #if 1
>        // template meta-programming implementation here
>   #else
>        // standard for loops implementation here
>   #endif
>
>
> which obviously was intended to be used temporarily for local debugging,
> but that unfortunately remained with the code...   for way too long.
>
>
> The consequence is that the "else" portion has not been tested recently,
> and therefore, it is broken, as probably illustrated by:
>
> http://www.cdash.org/CDash/testSummary.php?project=2&name=itkOrientedImage2DTest01&date=2008-10-19
>
>
> In addition, the "1" in the #if doesn't provide any indication
> of what the "if" is about. Which is an unfortunate lack of
> documentation for future maintenance of the code.
>
>
> Recent tests on gcc with -O3 illustrated that the compiler is
> actually able to unroll the for-loops and that the code with
> template meta-programming wasn't necessarily faster.
>
>
> By reconnecting this options to a CMake flag we can now:
>
>
>   1) Make sure that the "else" can be easily tested,
>      (maybe nightly, at least in some machines)
>
>   2) Re-evaluate the performance comparison of both
>      implementations.
>
>
>
> Your point of backward compatibility is well taken,
> therefore I'm turning the CMake flag:
>
>   ITK_USE_TEMPLATE_META_PROGRAMMING_LOOP_UNROLLING
>
> to be ON by default.
>
> http://www.itk.org/cgi-bin/viewcvs.cgi/CMakeLists.txt?root=Insight&view=log
> http://www.itk.org/cgi-bin/viewcvs.cgi/CMakeLists.txt?root=Insight&r1=1.318&r2=1.319
>
>
> This should bring the code back to behave just as when
> the  "#if 1"   was in place.
>
>
> We can now use the following Nightly tests:
>
>   itkOrientedImageProfileTest1.cxx
>   itkOrientedImageProfileTest2.cxx
>   itkOrientedImageProfileTest3.cxx
>
> to verify the computation time performance of the methods:
>
>   TransformIndexToPhysicalPoint()
>   TransformPhysicalPointToIndex(_
>
> in the
>
>   a) itk::OrientedImage  (test 1)
>   b) itk::Image          (test 2)
>   c) itk::VectorImage    (test 3)
>
>
> After running these profiling tests in multiple platforms
> we will be in a better informed position to decide whether
> we want to keep the option, or simply stick to one of the
> two implementations and remove the alternative code.
>
>
> My suggestion will be to base our test only on builds
> that use "Release" compilation mode, although one could
> argue that it may be desirable to have an option that could
> make Debug builds run faster...
>
>
> Thanks for any suggestions,
>
>
>     Luis
>
>
> ---------------------
> Bill Lorensen wrote:
>>
>> Luis,
>>
>> I noticed that the image helper methods are now enclosed within the
>> ITK_USE_TEMPLATE_META_PROGRAMMING_LOOP_UNROLLING  #ifdef. I suggest
>> that we do away with this cmake option. The image helper methods have
>> been the default behavior for oriented imaged from the beginning. I
>> see no reason to make that code optional now that we're also using it
>> in itkImage.
>>
>> Bill
>>
>