[Insight-developers] Performance Improvements for DiscreteGaussianImageFilter

Tue Jun 1 16:20:56 EDT 2010

Dear Brad,

You may consider also modifying the class

itk::DiscreteGaussianDerivativeImageFilter

which I contributed through IJ and is currently in Code/Review. I focused
mainly on the development of the derivative and Hessian operators, which
allowed their use as image functions, and basically copied the use of the
operators in an image filter (for now only for the Derivative operator) by
copying/adapting the corresponding implementation of
itk::DiscreteGaussianImageFilter. Thus, performance improvements should be
similar by modifying this class too. As I say the code is very very similar
(one could even consider inheritance in the future to avoid duplication).

My knowledge of the intrincacies of the streaming functionalities is
limited. In any case it seems to be a nice performance improvement, since
the performance gain increases with the kernel size (which is proportional
to the variance with the same parameters) and discrete Gaussian operators
have the problem that computation time increases with this variance. On the
other hand they allow local computations of Gaussians via image functions
which was actually my main interest :)

Regards

Ivan

2010/6/1 Bradley Lowekamp <blowekamp at mail.nih.gov>

> Hello,
>
> http://www.itk.org/Bug/view.php?id=10785
>
> I have noticed a few oddities in the way that DiscreteGaussianImageFilter
> is implement which impact it's performance. Unlike The recursive version,
> this filter is streamable, so I find it useful in many situations.
>
> What is unique about this filter is that internally it utilized
> StreamingImageFilter, to stream it's internal composite pipeline. It sets
> the number of streaming regions to ImageDimensions^2, with out any user
> control. Due to the way pipeline is implemented, it can result is extra
> computations in the overlap regions, and reduced multi-threaded CPU
> utilization due to excessive number of internal filters being executed. The
> can be particularly bad when the pipeline is already streaming, and the
> filter is also "sub-streaming".  I propose the following improvements:
>
>
> 1) Enable user settable number of division with an IVAR:
>
>   /** \brief Set/Get number of pieces to divide the input on the
>   * internal composite pipeline. The upstream pipeline will not be
>   * effected.
>   *
>   * The default value is $ImageDimension^2$.
>   *
>   * This parameter was introduced to reduce the size of images
>   * internally, at the cost of performance.
>    */
>   itkSetMacro(InternalNumberOfStreamDivisions,unsigned int);
>   itkGetConstReferenceMacro(InternalNumberOfStreamDivisions,unsigned int);
>
> 2) Reverse the internal pipeline. Reverse 1-D convolutions from XYZ to ZYX.
> Because the streamer slices in the z direction, the remaining dimensions are
> already expanded and will not grow in the GenerateInputRequestedRegion phase
> of the pipeline. Therefore to minimize the computation the slice direction Z
> should be done first. This results in no recompilation or overlapping
> regions from the convolution filters.
>
>
> Depending on the size of the image, and the amount of cache the performance
> gain by this improvement is 20-60%. I have attached the enhanced filter and
> a test driver.
>
> creating image with size: [128, 128, 128]
> randomizing image
> =======================================
> Executed base with variance 0.0625 in 0.0746817s
> Executed base with variance 0.25 in 0.105442s
> Executed base with variance 1 in 0.122377s
> Executed base with variance 4 in 0.202058s
> Executed base with variance 16 in 0.517367s
> Executed base with variance 64 in 2.2312s
> =======================================
> Executed improved ordered with variance 0.0625 in 0.0739802s
> speedup 0.00939382
> Executed improved ordered with variance 0.25 in 0.102979s
> speedup 0.0233524
> Executed improved ordered with variance 1 in 0.117431s
> speedup 0.0404161
> Executed improved ordered with variance 4 in 0.176935s
> speedup 0.124332
> Executed improved ordered with variance 16 in 0.349574s
> speedup 0.324321
> Executed improved ordered with variance 64 in 0.969516s
> speedup 0.565473
> =======================================
> Executed with 1 stream division with variance 0.0625 in 0.0664336s
> speedup 0.110444
> Executed with 1 stream division with variance 0.25 in 0.0864969s
> speedup 0.179669
> Executed with 1 stream division with variance 1 in 0.110848s
> speedup 0.0942072
> Executed with 1 stream division with variance 4 in 0.168331s
> speedup 0.166916
> Executed with 1 stream division with variance 16 in 0.394602s
> speedup 0.237289
> Executed with 1 stream division with variance 64 in 1.15303s
> speedup 0.483224
> =======================================
> Executed recursive gaussian with variance 0.0625 in 0.0472106s
> speedup 0.367842
> Executed recursive gaussian with variance 0.25 in 0.0468103s
> speedup 0.556054
> Executed recursive gaussian with variance 1 in 0.0453837s
> speedup 0.629148
> Executed recursive gaussian with variance 4 in 0.0511486s
> speedup 0.746861
> Executed recursive gaussian with variance 16 in 0.0571482s
> speedup 0.88954
> Executed recursive gaussian with variance 64 in 0.0671696s
> speedup 0.969895
>
> Clearly the SmoothingRecursiveGaussianImageFilter is the performance winner
> here (the synthetic image may be giving accurate results), and future
> efforts should really include making this filter streamable.
>
> However, I still think these improvements are important to contribute. As I
> believe they are fully backwards compatible, I will try to commit them this
> week if no one sees any issues.
>
>
> Thanks,
> Brad
>
>
>
>
>
> ========================================================
>
> Bradley Lowekamp
>
> Lockheed Martin Contractor for
>
> Office of High Performance Computing and Communications
>
> National Library of Medicine
>
> blowekamp at mail.nih.gov
>
>
>
> _______________________________________________
> 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://kitware.com/products/protraining.html
>
> 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-developers
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/mailman/private/insight-developers/attachments/20100601/2d65f3ef/attachment.htm>