Custom border extrapolation of ShapedImageNeighborhoodRange by ImageNeighborhoodPixelAccessPolicy

[matt.mccormick]

@Niels_Dekker

Using Visual C++ 2017 64-bit Release, I observed that the range-based neighborhood iteration implemented by http://review.source.kitware.com/#/c/23795/7 can be more than twice as fast as the “old-school” iteration

For the record, it was only after I wrote those words, that I found a way to significantly improve the performance of the “old-school” iteration (using Visual C++ 2017 Release): PERF: Made ConstNeighborhoodIterator::GetPixel(i) much faster

@Niels_Dekker Well done! To add more data to the Visual C++ 2017 Release results, here are some metrics created on Linux with GCC 7.3.0 and ICC (the Intel Compiler) 19.0.0.120. These results are created for this ultrasound application and your Gist.

They are very interesting! All we can conclude is that neither implementation is consistently faster across compilers!

That said, the potential of these new image range classes to be used with STL algorithms is quite compelling because of potential future benefits:

• C++17 Parallel STL execution policies
• Intel parallelstl TBB-backed implementation of the above
• OpenCL / SYCL GPU / FPGA implementation of parallel STL via ComputeCPP or triSYCL
• AMD support
• Potential NVidia support

Granted that we would have to figure out a good way to control data synchronization in the GPU case.

@phcerdan

The usage:

// The equivalent in ITK 5.0 + this patch +  this
// Using itk::Experimental::ShapedImageNeighborhoodRange:
using RangeType = ShapedImageNeighborhoodRange< ImageType,
ConstantBoundaryImageNeighborhoodPixelAccessPolicy<ImageType> >;

RangeType range{ *image, location, offsets}; // same constructor as before
range->GetReferencePixelAccessPolicy()->SetConstant(paddingValue); // we set the value at run time on the policy

+1 to this API – setting the constant value separately from the range initialization is easier to follow, in my opinion. That said, this is not implemented, and maybe we could implement it in the future if it is really compelling. As an itk::Experimental namespaced feature, we should keep experimenting with nice things!

Let’s merge the patch, which @hjmjohnson has kindly migrated to GitHub!