[Insight-developers] OptBSplineInterpolateImageFunction

[M.Staring at lumc.nl]

Hi,

Just to share timing results on my platform, so that we can better
estimate impact on performance on several platforms.

On my platform (Windows XP SP3, using Visual C++ 2008 release mode, 3rd
order 
> Bspline, 3d 100x100x100 image, ITK 3.14, Intel Core 2 quad @ 2.4 GHz
Q6600) I get:

The elapsed time for old way: 3297
The elapsed time for fast way: 2437
The elapsed time for opt way: 2141 

So a bit more in favour of opt, but your code indeed looks much more
simple.

Marius

> -----Original Message-----
> From: insight-developers-bounces at itk.org 
> [mailto:insight-developers-bounces at itk.org] On Behalf Of Stefan Klein
> Sent: Thursday, August 27, 2009 5:06 PM
> To: insight-developers at itk.org
> Subject: [Insight-developers] OptBSplineInterpolateImageFunction
> 
> Hi,
> 
> I wrote an alternative to the itkOptBSplineInterpolateImageFunction, 
> which is nearly as fast, but requires less code and is easier 
> to use (in 
> my opinion).
> 
> The OptBSplineInterpolateImageFunction adds methods that take the 
> threadId as argument. For example ::EvaluateAtIndex( index, 
> threadId). 
> This allows some dynamic memory allocations to be avoided and gives a 
> great speedup.
> 
> An alternative way of avoiding the memory allocations is to declare a 
> C-array (of fixed size, so allocated on the stack) within the 
> EvaluateAtIndex method, and use the vnl_matrix_ref class to 
> instantiate 
> a matrix that uses the statically allocated memory (instead of 
> allocating memory on the heap).
> 
> 1 million calls to EvaluateAtIndex give the following result:
> The elapsed time for old way: 2953ms
> The elapsed time for fast way: 2219ms
> The elapsed time for opt way: 2110ms
> 
> old = non-optimized itkBSplineInterpolateImageFunction
> fast = my proposed implementation
> opt = the itkOptBSplineInterpolateImageFunction.
> (tested on Windows XP, using Visual C++ 2008 release mode, 3rd order 
> Bspline, 3d 100x100x100 image, ITK 3.14)
> 
> In the attachment you can find the test program and the classes. My 
> proposed implementation is the 
> itkBSplineInterpolateImageFunctionFast.h. 
> I renamed the Opt-classes in the testing code, to make it easier to 
> compare them, but I didn't change the functional code.
> 
> The opt version is a bit faster, but requires more lines of code (and 
> requires changes to all code that uses it, in order to 
> benefit from the 
> fast implementation). For the EvaluateDerivativeAtIndex method the 
> timing results are similar. All interpolated values are 
> identical in the 
> three implementations (old,fast,opt).
> 
> What do you think about it?
> 
> Kind regards,
> Stefan
> 
> 
> 
> 
> No virus found in this incoming message.
> Checked by AVG - www.avg.com 
> Version: 8.5.409 / Virus Database: 270.13.69/2328 - Release 
> Date: 08/26/09 12:16:00
>